AI Citations

OpenAI uses a search crawlerA crawler — also called a spider or bot — is an automated program that fetches web pages, extracts their links, and queues new URLs to visit. Search engines use crawlers to discover and download content for their index. distinct from its training crawlerAI crawlers are bots from AI companies that fetch web pages to train language models, build AI-search indexes, or answer live user questions. They come in three categories, each with its own user-agent tokens and its own robots.txt controls. to surface and link to websites in ChatGPT search. Evidence for this claim OpenAI identifies OAI-SearchBot as the crawler used to link to and surface websites in ChatGPT search results, separately from GPTBot training controls. Scope: OpenAI-declared crawler behavior; selection and citation are not guaranteed. Confidence: high · Verified: 2026-07-14 OpenAI: Crawlers Google likewise describes supporting links as part of its AI searchAI search uses large language models and retrieval-augmented generation (RAG) to synthesize an answer from multiple sources rather than returning a ranked list of links. Examples include Google AI Overviews, ChatGPT Search, and Perplexity. experiences. Evidence for this claim Google says AI Overviews and AI Mode show supporting links and can use query fan-out to find a wider and more diverse set of relevant pages. Scope: Google Search AI features; link selection is system-dependent. Confidence: high · Verified: 2026-07-14 Google: AI features and your website

TL;DR — An AI citation is the little source link an AI answer shows next to what it wrote — the clickable reference that says “this came from that page.” Getting cited is not the same as ranking #1, and it’s not the same as having lots of backlinks — and it’s not proof the source actually backs up what the AI said, which is a separate thing to check. It runs on the AI reading real web pages and deciding which ones to credit. To be in the running, your page has to be findable and readable — and the internet talking about you helps more than most people expect.

What an AI citation is

An AI citation is the visible source link an AI answer shows next to what it wrote. That’s the whole definition — but a citation showing up doesn’t mean the cited page actually backs up the claim sitting next to it. Those are two separate things to check: is the source shown, and does the source actually support what the AI said.

When you ask ChatGPT, Perplexity, Google’s AI OverviewsAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index., or Microsoft Copilot a question, you get a written answer, and some responses include source links. Those links are citations: the pages the product credits for that response. But a missing link is not automatically a problem with your page. Google says Gemini Apps do not include sources on every response, Evidence for this claim Gemini Apps may provide sources or related links, but not every response includes them; no Sources button means Gemini provided no links for that response. Scope: Gemini Apps source-link presentation; absence of links does not by itself establish retrieval failure or a publisher defect. Confidence: high · Verified: 2026-07-19 Google Gemini Apps: View related sources and Microsoft says Copilot Voice currently does not run web search, so voice responses have no citations by design. Evidence for this claim Microsoft's consumer Copilot transparency note says web-grounded text responses include citations, while voice prompts do not trigger web search and therefore have no citations. Scope: Consumer Microsoft Copilot Voice mode; Microsoft 365 Copilot and text/web-grounded modes have different behavior. Confidence: high · Verified: 2026-07-19 Microsoft Copilot: Transparency note

Think of it like a school essay. The AI writes the answer in its own words, but it shows its sources so you can check them (and click through). Being one of those sources is what “getting cited by AI” means — proving the source actually backs the claim next to it is a second, separate check.

A citation isn’t proof the source is right

This is the part beginners skip and it causes real mistakes: a citation is an attribution, not a fact-check. Research on early generative-search systems foundA 302 (\"Found\") is a temporary redirect: it forwards users to a new URL while telling search engines the original URL should stay in the index. It's a weak canonicalization signal, not the zero-equity dead end of SEO folklore. real gaps between “a citation is shown” and “the citation actually supports the sentence next to it” — some cited pages didn’t fully back up the claim they were attached to. If you’re using a cited page to make a decision, read the page against the specific claim before you trust it.

Retrieved, mentioned, cited — three different things

Here’s the part almost everyone blurs together. There are three separate things an AI can do with your content, and they don’t always happen together:

1. Retrieved — the AI read your page behind the scenes to help write its answer. You can only see this in your server logsLog file analysis is reading a web server's raw access logs to see exactly which URLs search engine crawlers actually requested, when, how often, and what status code they got. Unlike crawl tools or Search Console, logs are the unsampled, ground-truth record of what really happened..
2. Mentioned — your brand name shows up in the answer text, but with no link.
3. Cited — your actual URL is shown as a source you can click.

Your content can be read without being named, and named without being linked. Most AI-visibility tools only see the last one (citations), which is why so much of your real influence stays invisible.

Why it’s not just “rank #1 and you’re cited”

The intuitive assumption — “if I’m the top Google result, the AI will cite me” — is increasingly wrong. AI answers often pull sources from a bunch of related behind-the-scenes searches, not just the top results for your exact question. A page can be cited without ranking in the top 10, and a #1 page can be skipped.

There’s no separate “AI indexStoring a crawled page in the search index so it can appear in results. Crawled is not the same as indexed — Google selects what to keep, and indexing isn't guaranteed.” for Google, either — its AI OverviewsAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index. use the same Search index as normal results. So the basics still gate everything: if your page can’t be crawled and indexed, it can’t be cited. Beyond that, being talked about on other sites (Reddit, YouTube, industry blogs) turns out to matter a lot — often more than your own site does.

The catch: cited doesn’t always mean traffic

Even when you get cited, most people don’t click. A lot of AI answers give people what they needed without a visit. So “get cited” is a real goal, but it’s an awareness and authority signal first, and a traffic driver a distant second.

Want the real mechanics — how RAGRAG is the retrieve-then-generate pattern behind AI search: the system retrieves relevant passages from an external index at query time, injects them into the model's context, and generates an answer grounded in those sources — without changing the model's weights. picks sources, the log-file bots that reveal retrieval, the citation-rate differences across platforms, and the actual optimization levers? Switch to the Advanced tab.

For OpenAI, OAI-SearchBot’s declared purpose is search visibility, while GPTBot has a separate training control. Evidence for this claim OpenAI identifies OAI-SearchBot as the crawler used to link to and surface websites in ChatGPT search results, separately from GPTBot training controls. Scope: OpenAI-declared crawler behavior; selection and citation are not guaranteed. Confidence: high · Verified: 2026-07-14 OpenAI: Crawlers For Google, supporting-link selection can draw on query fan-outQuery fan-out is the technique where an AI search system breaks a single user question into multiple related sub-queries, runs those searches concurrently, and synthesizes the retrieved results into one answer. Google confirms AI Overviews and AI Mode 'may use a query fan-out technique' issuing multiple related searches across subtopics. beyond a single result set. Evidence for this claim Google says AI Overviews and AI Mode show supporting links and can use query fan-out to find a wider and more diverse set of relevant pages. Scope: Google Search AI features; link selection is system-dependent. Confidence: high · Verified: 2026-07-14 Google: AI features and your website And a citation appearing at all is never proof that the cited page supports the adjacent claim — that’s a distinct check layered on top of attribution. Evidence for this claim Citation presence alone is not proof the cited source supports the adjacent generated statement; a study of four 2023 generative search systems found meaningful gaps between citation recall and citation precision. Scope: Findings from evaluated 2023 generative search systems (Bing Chat, NeevaAI, Perplexity, YouChat); supports the general citation-verifiability distinction, not current-system rates. Confidence: high · Verified: 2026-07-15 Liu, Zhang & Liang: Evaluating Verifiability in Generative Search Engines (EMNLP 2023)

TL;DR — An AI citation is the attributed source shown next to AI-generated text, selected at generation time from what the system retrieved via RAGRAG is the retrieve-then-generate pattern behind AI search: the system retrieves relevant passages from an external index at query time, injects them into the model's context, and generates an answer grounded in those sources — without changing the model's weights. — not ranked into a list. It’s the last of three states: retrieved (a logged request for your page; not proof the content reached the model’s context), mentioned (named in the text, no link), and cited (URL attributed). Google grounds AI Overviews in its core indexStoring a crawled page in the search index so it can appear in results. Crawled is not the same as indexed — Google selects what to keep, and indexing isn't guaranteed., so a page must be indexed and snippet-eligible to be cited — but citation ≠ ranking (Ahrefs research points to the top-10 share falling sharply, roughly 76% toward roughly 38%, between mid-2025 and early 2026, though the exact methodology comparison across the two studies is still unverified) and ≠ backlinks (branded web mentions correlate ~0.66 with AI visibility vs. ~0.33 for Domain Rating). Roughly 90% of citations about a brand come from third-party sites, and platforms cite wildly differently (Perplexity cites almost everything it retrieves; ChatGPT cites almost nothing). Even a citation usually produces no click — and a citation is never proof the source supports the claim next to it.

Citation is the last of three states, not the whole picture

Start with the distinction that fixes most measurement confusion. Your content can be in one, two, or all three of these states at once — and each requires a different tool to observe:

• Retrieved. A log hit from an AI crawlerAI crawlers are bots from AI companies that fetch web pages to train language models, build AI-search indexes, or answer live user questions. They come in three categories, each with its own user-agent tokens and its own robots.txt controls. or live-inference botA crawler — also called a spider or bot — is an automated program that fetches web pages, extracts their links, and queues new URLs to visit. Search engines use crawlers to discover and download content for their index. shows a request was made to your page while the system processed a query. That’s real evidence of a fetch — but a request in your logs is not proof the content was pulled into the model’s context or used at all; logs show what was requested, not what was selected. This is detectable only in server logsLog file analysis is reading a web server's raw access logs to see exactly which URLs search engine crawlers actually requested, when, how often, and what status code they got. Unlike crawl tools or Search Console, logs are the unsampled, ground-truth record of what really happened. (AI crawlerA crawler — also called a spider or bot — is an automated program that fetches web pages, extracts their links, and queues new URLs to visit. Search engines use crawlers to discover and download content for their index. user-agentsA user agent is the HTTP request header a client (browser, crawler, or bot) sends to identify itself. For crawlers, a short user-agent token — a substring of that string — is what robots.txt rules actually target.), and it’s invisible to any tool that works purely by analyzing AI responses.
• Mentioned. Your brand, product, or name appears in the generated text — with no hyperlink. The AI “knows about you” from training data or from retrieved content and names you without attribution. Measurable via response-analysis tools like Ahrefs Brand Radar.
• Cited. Your page’s URL is shown as an attributed source — a clickable link, a numbered reference, or a named source box. This requires the AI to have both retrieved your content and decided it was worth attributing.

The gaps between them are the whole story. Content can be retrieved without being mentioned; mentioned without being cited; and — the inverse “ghost citation” problem — cited without your brand name ever appearing in the text. And a request observed in your logs is a distinct, weaker claim than “selected into the answer” — treat retrieval as evidence of a fetch, not proof of use. (The dedicated retrieved / mentioned / cited framing is worth its own treatment; this article is the citation-mechanics end of it.)

How the citation actually gets made: RAG over a live index

Citations are an output of Retrieval-Augmented Generation — the retrieve-then- generate pattern behind every major AI answer engine. Google’s own definition ties it directly to groundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it.: RAG is “a technique (also known as groundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it.) used to improve the quality, accuracy, and freshness of AI responses by relying on our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index.” (For the retrieval pipeline itself — chunkingChunking is splitting a document into smaller passages so AI systems can embed, index, and retrieve the single most relevant piece — not the whole page — in response to a query. It's a foundational step in RAG pipelines and the conceptual cousin of Google's passage ranking., embeddingsEmbeddings are dense numerical vectors — lists of floating-point numbers — that represent the meaning of text in a high-dimensional space. Semantically similar content lands close together, so search and AI systems can match by meaning, not just keywords., vector searchVector search finds content by comparing the meaning of a query against stored content as numerical vectors, retrieving the closest ones in a high-dimensional embedding space. At scale it uses approximate nearest neighbor (ANN) algorithms — not exact comparison — to search billions of vectors in milliseconds., re-ranking — see the sibling deep dives; here I care about what happens at the attribution step.)

Two consequences follow directly from that definition, and they’re the ones Google states outright:

• There is no separate AI index. AI OverviewsAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index. retrieve from the same core Search index as normal results, so “a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements. There are no additional technical requirements.” Crawlable, indexable, snippet-eligible is the citation floor. Block GooglebotGooglebot is Google's web crawler — the software that fetches pages so Google can index and rank them. It comes in two variants, Googlebot Smartphone (primary, under mobile-first indexing) and Googlebot Desktop, and runs an evergreen Chromium renderer. or return errors and you’re removed from eligibility entirely.
• Retrieval is broader than the single query. Google runs query fan-outQuery fan-out is the technique where an AI search system breaks a single user question into multiple related sub-queries, runs those searches concurrently, and synthesizes the retrieved results into one answer. Google confirms AI Overviews and AI Mode 'may use a query fan-out technique' issuing multiple related searches across subtopics.: “Both AI Overviews and AI Mode may use a ‘query fan-out’ technique — issuing multiple related searches across subtopics and data sources — to develop a response.” Pages get retrieved for sub-queries that never appear in the visible answer. Those are the “dark” retrievals — content that contributed without attribution.

At the API level the architecture makes the retrieved-but-not-cited gap explicit: Google’s grounded generation returns the model’s text with inline citation annotations, each linking a specific text span to a source URL — but the full list of retrieved sources isn’t exposed to the end user. You see what got used, not what got read.

Why citation is not ranking (and the number that proves it)

One of the most useful directional findings for anyone doing this work: being cited by AI appears to be decoupling from ranking #1 over time. In Ahrefs’ research (the July 2025 study, then a March 2026 update run by Xibeijia Guan), the share of AI Overview citations that came from a top-10 organic position is reported to have fallen from roughly 76% in mid-2025 to about 38% by early 2026 — with BrightEdge’s independent number even lower. I want to flag this plainly: I haven’t been able to independently re-verify the exact methodology and denominator behind that 76%-to-38% comparison, so treat the direction (a declining top-10 share) as the takeaway, not the precise percentages, until that check is done. Google’s own explanation for the broader trend is that query fan-out is doing more of the sourcing: AI Overviews increasingly pull from the SERPs of sub-queries rather than the original query’s top 10 — and that mechanism is confirmed directly in Google’s own documentation, independent of the exact stat above.

Two practical reads:

• Ranking #1 for your head query no longer guarantees a citation, and not ranking in the top 10 no longer excludes you. YouTube is a striking example — it’s the most-cited domain in AI Overviews and a large share of the YouTube URLs cited don’t rank in Google’s top 100 for the query at all. Video gets pulled in largely independent of classic ranking.
• Citation eligibility is still gated by classic SEO (indexed + snippet-eligible) even as citation selection drifts away from rank position. Both things are true at once, which is why “AI killed SEO” and “AI is just SEO” are both wrong.

Why citation is not backlinks either

The other reflex — “more backlinks, more citations” — also underperforms the data. Ahrefs’ 75,000-brand study by Louise Linehan and Xibeijia Guan found that the strongest predictor of appearing in AI Overviews wasn’t Domain Rating or backlinks; it was branded web mentions, correlating around 0.664, versus roughly 0.326 for Domain Rating and lower still for backlink counts. In a separate analysis I ran, mentions on heavily linked pages also correlated with AI mentions.

That reframes the whole game. It’s not that authority stopped mattering — it’s that being talked about is a better proxy for what AI systems reward than being linked to. McKinsey’s finding sharpens it: a brand’s own website accounts for only about 5–10% of the sources AI platforms reference when discussing that brand. Roughly 90% of the citations about you live on third-party sites — publishers, review platforms, UGC, YouTube, Reddit. As I put it in my Ahrefs Evolve talk: the job used to be “optimize your site,” and now it’s “optimize how the internet talks about you.” Sixteen pages on Zapier’s site that mention Ahrefs were enough to get Ahrefs cited in over 1,400 AI responses — a level of leverage no amount of on-site tweaking matches.

Platforms cite completely differently

“Getting cited by AI” isn’t one behavior — the same page can be retrieved on two platforms with opposite attribution outcomes:

• Perplexity runs a real-time web search for essentially every query and is built to be transparent about sources — it averages the most citations per response of any major platform and cites nearly everything it retrieves. It’s the least likely to “mention without citing.”
• ChatGPT answers a great deal from training data with no retrieval at all, and when it does retrieve (via its Bing-powered search layer) it cites only a small fraction of the pages it fetched. Its overall citation rate is the lowest of the major platforms. This is the purest “mentioned but not cited” surface.
• Google AI Overviews / AI Mode attach citations to specific text segments, source from the core index plus fan-out, and — as above — increasingly cite outside the top 10.
• Microsoft Copilot exposes something the others don’t: grounding queries, the actual phrases it used to retrieve content that was referenced. That’s the closest any official tool comes to showing the retrieval layer.

The takeaway isn’t to chase one platform’s quirks — it’s that citation is probabilistic and inconsistent. AI responses are volatile; the same prompt run repeatedly produces different brand lists and different sources, so a single check tells you almost nothing.

There are also documented product and mode boundaries that a monitor should store as distinct outcomes instead of calling every missing link a publisher issue:

• OpenAI’s web-search API: the sources field is the complete set of URLs the model consulted, while inline citations are a smaller selected set. A source can therefore be retrieved without being cited, and the API docs require displayed inline citations to be visible and clickable. This is an API contract, not proof that the ChatGPT consumer product exposes the same fields. Evidence for this claim In OpenAI API web search, the sources field lists all consulted URLs while inline citations show only selected relevant references; source count can exceed citation count, and search_context_size guarantees neither. Scope: OpenAI API web search response fields and configuration; not a claim about undisclosed retrieval or ChatGPT consumer-product presentation. Confidence: high · Verified: 2026-07-19 OpenAI API: Web search — Sources and search context size Evidence for this claim OpenAI requires inline citations to be clearly visible and clickable when an application displays web-search results or information from those results to end users. Scope: OpenAI API web search display contract; this does not assert identical fields or presentation in the ChatGPT consumer product. Confidence: high · Verified: 2026-07-19 OpenAI API: Web search — Output and citations
• Microsoft Copilot: consumer Voice currently does not trigger web search, so no citation is expected. Evidence for this claim Microsoft's consumer Copilot transparency note says web-grounded text responses include citations, while voice prompts do not trigger web search and therefore have no citations. Scope: Consumer Microsoft Copilot Voice mode; Microsoft 365 Copilot and text/web-grounded modes have different behavior. Confidence: high · Verified: 2026-07-19 Microsoft Copilot: Transparency note In Microsoft 365 Copilot, when web search does run, the Sources view shows both the sources and the exact short query sent to Bing; that derived query usually uses only a few words from the prompt or file. Evidence for this claim Microsoft 365 Copilot's Sources view exposes the exact query sent to Bing and the sources used; the query generally uses a few prompt or uploaded-file words rather than the full input. Scope: Microsoft 365 Copilot Chat and agents when web search runs; organization policy can disable web search. Confidence: high · Verified: 2026-07-19 Microsoft 365 Copilot: How web search works
• Gemini Apps: Google explicitly says not every response includes related links or sources. No Sources button means no links were provided for that response; it does not by itself prove that retrieval failed or that the publisher blocked anything. Evidence for this claim Gemini Apps may provide sources or related links, but not every response includes them; no Sources button means Gemini provided no links for that response. Scope: Gemini Apps source-link presentation; absence of links does not by itself establish retrieval failure or a publisher defect. Confidence: high · Verified: 2026-07-19 Google Gemini Apps: View related sources

The uncomfortable truth: cited ≠ traffic

Finally, temper the goal. Being cited is an awareness and authority signal far more than a traffic channel right now. Independent research (Pew) found that when an AI summary appears, only a small single-digit percentage of users click any link at all, and only about 1% click a cited source within the summary — with the large majority of AI searchAI search uses large language models and retrieval-augmented generation (RAG) to synthesize an answer from multiple sources rather than returning a ranked list of links. Examples include Google AI Overviews, ChatGPT Search, and Perplexity. sessions ending with no site visit. Optimize for citations because they build brand presence and authority in the answers people read — not because you should expect a click for each one.

Where this fits: citation is the attribution end of grounding and RAG, it runs on retrieval (chunking, embeddings, vector search, passage rankingPassage ranking is a Google AI system that scores individual sections ('passages') of a page so a single page can earn multiple relevance scores for different queries. Google still indexes whole pages — only the ranking changed.) upstream, and it lives next to the freshness and hallucination dynamics of AI answers. The mechanics are shared across this cluster; what’s specific to citations is the three-state distinction, the ranking/backlink decoupling, and the third-party-mention leverage.

AI summary

A condensed take on the Advanced version:

• An AI citation = the attributed source shown next to AI-generated text, chosen at generation time from what the system retrieved via RAGRAG is the retrieve-then-generate pattern behind AI search: the system retrieves relevant passages from an external index at query time, injects them into the model's context, and generates an answer grounded in those sources — without changing the model's weights. — not ranked into a blue-link list.
• Three states, three tools: retrieved (a logged request for your page, not proof the content reached the model’s context; visible only in server logsLog file analysis is reading a web server's raw access logs to see exactly which URLs search engine crawlers actually requested, when, how often, and what status code they got. Unlike crawl tools or Search Console, logs are the unsampled, ground-truth record of what really happened.), mentioned (named in text, no link; response-analysis tools), cited (URL attributed; citation-tracking tools). Content can be retrieved without being mentioned, mentioned without being cited, and even cited without being named (“ghost citations”).
• Runs on RAG over the core indexStoring a crawled page in the search index so it can appear in results. Crawled is not the same as indexed — Google selects what to keep, and indexing isn't guaranteed.. Google: no separate AI index; a page must be indexed and snippet-eligible to be cited — that’s the floor. Query fan-outQuery fan-out is the technique where an AI search system breaks a single user question into multiple related sub-queries, runs those searches concurrently, and synthesizes the retrieved results into one answer. Google confirms AI Overviews and AI Mode 'may use a query fan-out technique' issuing multiple related searches across subtopics. retrieves for sub-queries, creating “dark” (uncited) retrievals.
• Citation ≠ ranking. Ahrefs research points to the top-10 organic share of AI Overview citations falling sharply between mid-2025 and early 2026 (roughly 76% toward roughly 38%) — the exact methodology behind that comparison is still unverified, so treat the direction, not the precise numbers, as settled; YouTube is the most-cited domain and often isn’t in the top 100 for the query.
• Citation ≠ backlinks. Branded web mentions correlate ~0.66 with AI visibility vs. ~0.33 for Domain Rating; ~90% of citations about a brand come from third-party sites (McKinsey: own site = 5–10%).
• Citation ≠ support. A cited page isn’t automatically proven to back up the claim next to it — research on 2023 generative-search systems foundA 302 (\"Found\") is a temporary redirect: it forwards users to a new URL while telling search engines the original URL should stay in the index. It's a weak canonicalization signal, not the zero-equity dead end of SEO folklore. real gaps between citation presence and citation support. Check the claim against the page, not just whether a link is shown.
• Platforms differ hugely: Perplexity cites nearly everything it retrieves; ChatGPT cites almost nothing and often answers from training data. Copilot uniquely exposes groundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it. queries.
• Cited ≠ traffic. Most AI sessions end without a click; ~1% click cited sources within summaries (Pew). Treat citation as an authority/awareness signal.
• Strategic shift: from “optimize your site” to “optimize how the internet talks about you."

"I want more AI citations — where do I start?”

Citation problems live at different stages. Diagnose which stage is failing before you change anything — the fix for each is completely different.

Step 1 — Can the AI even reach you? Check your server logsLog file analysis is reading a web server's raw access logs to see exactly which URLs search engine crawlers actually requested, when, how often, and what status code they got. Unlike crawl tools or Search Console, logs are the unsampled, ground-truth record of what really happened. and robots rules.

• Not seeing AI crawlerAI crawlers are bots from AI companies that fetch web pages to train language models, build AI-search indexes, or answer live user questions. They come in three categories, each with its own user-agent tokens and its own robots.txt controls. hits (Googlebot-Extended, OAI-SearchBot, PerplexityBot, the -User inference botsA crawler — also called a spider or bot — is an automated program that fetches web pages, extracts their links, and queues new URLs to visit. Search engines use crawlers to discover and download content for their index.)? → You may be blocking them, or you’re not retrievable at all. Fix crawlabilityCrawlability is how well search engine crawlers can discover, access, and fetch a site's pages. A crawlability issue is any technical condition — blocked access, broken links, server failures, or bloated URL inventory — that stops pages from reaching the index. first — nothing downstream matters until you can be fetched.
• Seeing -User bots hit error responses? → Every error to a live-inference bot is a missed citation in a real conversation. Fix those responses.

Step 2 — Are you retrieved but not cited? Logs show retrieval, but you’re not appearing as a source.

• → This is a quality/structure/completeness problem, not an access one. Make the answer self-contained and clearly the best source for the sub-question. Being retrieved is necessary but far from sufficient (ChatGPT cites only ~15% of what it retrieves).

Step 3 — Are you mentioned but not cited? Your brand name appears in answers, but with no link to you.

• → The AI knows you but doesn’t attribute you. Strengthen direct attribution signals and make your own page the obvious canonical source for the specific claim. Also expect some of this to be structural (ChatGPT-style “mention without cite”).

Step 4 — Are you barely present at all (not even mentioned)?

• → This is a reputation problem, and it’s mostly off-site. ~90% of citations about a brand come from third-party sites. Get talked about on the sources AI actually pulls from (industry publishers, Reddit, YouTube, review sites). Optimize how the internet talks about you, not just your own pages.

Step 5 — Cited but getting no traffic?

• → Expected, for now. Most AI sessions end without a click. Shift the success metric to brand presence in answers (mention + citation together), not clicks — and make sure you’re named, not just linked, so users associate the answer with you.

Which measurement layer answers each question?

Official documentation

Primary-source documentation on how AI features source and cite content.

Google

• AI features and your website — eligibility for supporting links in AI OverviewsAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index./AI Mode, query fan-outQuery fan-out is the technique where an AI search system breaks a single user question into multiple related sub-queries, runs those searches concurrently, and synthesizes the retrieved results into one answer. Google confirms AI Overviews and AI Mode 'may use a query fan-out technique' issuing multiple related searches across subtopics., and how AI-feature traffic is reported in Search ConsoleA free Google service that reports how a site performs in Google Search and surfaces problems with how Google crawls, indexes, and serves it. It's first-party data straight from Google — but you don't need it to appear in results..
• Google’s Guide to Optimizing for Generative AI Features — the RAGRAG is the retrieve-then-generate pattern behind AI search: the system retrieves relevant passages from an external index at query time, injects them into the model's context, and generates an answer grounded in those sources — without changing the model's weights./groundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it. definition over the core indexStoring a crawled page in the search index so it can appear in results. Crawled is not the same as indexed — Google selects what to keep, and indexing isn't guaranteed., plus the “still SEO” framing and mythbusting (no llms.txtllms.txt is a proposed (not adopted) Markdown file at /llms.txt that gives AI systems a curated map of a site's most important pages. Proposed by Jeremy Howard in 2024, it's read mostly by coding agents like Claude Code — not search crawlers — and Google ignores it., no chunkingChunking is splitting a document into smaller passages so AI systems can embed, index, and retrieve the single most relevant piece — not the whole page — in response to a query. It's a foundational step in RAG pipelines and the conceptual cousin of Google's passage ranking. requirement, no special schema).
• AI Overviews and AI Mode in Search — confirms no additional requirements beyond standard indexingStoring a crawled page in the search index so it can appear in results. Crawled is not the same as indexed — Google selects what to keep, and indexing isn't guaranteed. and snippet eligibility.
• Grounding with Google Search (Gemini API) — the developer-level groundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it. flow: search execution, synthesis, and inline url_citation annotations linking text spans to source URLs.
• Search Generative AI performance reports in Search Console — the June 2026 report giving impressions-only visibility for AI features (no clicks/CTR/query data yet).

Microsoft / Bing

• Introducing AI Performance in Bing Webmaster Tools (public preview) — the first official platform tool to expose citation counts and grounding queries to webmasters.
• How web search works in Microsoft 365 Copilot — the Sources view, derived Bing query, and organization-level web-search control.
• Transparency Note for Microsoft Copilot — consumer text grounding and the current no-web-search/no-citation exception for voice prompts.

OpenAI / Gemini

• OpenAI API web search — visible/clickable citation display requirements and the distinction between all consulted sources and selected inline citations.
• View related sources from Gemini Apps — Google states that not every Gemini Apps response includes links or a Sources button.

Quotes from the source

On-the-record statements from Google and Microsoft. Each Google link is a deep link that jumps to the quoted passage on the source page.

Google — RAGRAG is the retrieve-then-generate pattern behind AI search: the system retrieves relevant passages from an external index at query time, injects them into the model's context, and generates an answer grounded in those sources — without changing the model's weights., groundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it., and the core indexStoring a crawled page in the search index so it can appear in results. Crawled is not the same as indexed — Google selects what to keep, and indexing isn't guaranteed.

• “A technique (also known as groundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it.) used to improve the quality, accuracy, and freshness of AI responses by relying on our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index.” — Google Search Central, AI optimization guide. Jump to quote
• “Our generative AI features on Google Search are rooted in our core Search ranking and quality systems.” — Google Search Central, AI optimization guide.

Google — what makes a page eligible to be cited

• “To be eligible to be shown as a supporting link in AI OverviewsAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index. or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements. There are no additional technical requirements.” — Google Search Central, AI features and your website. Jump to quote
• “There are no additional requirements to appear in AI OverviewsAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index. or AI Mode, nor other special optimizations necessary.” — Google Search Central, AI features and your website. Jump to quote

Google — query fan-outQuery fan-out is the technique where an AI search system breaks a single user question into multiple related sub-queries, runs those searches concurrently, and synthesizes the retrieved results into one answer. Google confirms AI Overviews and AI Mode 'may use a query fan-out technique' issuing multiple related searches across subtopics. (why “dark” retrievals exist)

• “Both AI Overviews and AI Mode may use a ‘query fan-outQuery fan-out is the technique where an AI search system breaks a single user question into multiple related sub-queries, runs those searches concurrently, and synthesizes the retrieved results into one answer. Google confirms AI Overviews and AI Mode 'may use a query fan-out technique' issuing multiple related searches across subtopics.’ technique — issuing multiple related searches across subtopics and data sources — to develop a response.” — Google Search Central, AI features and your website. Jump to quote
• “While responses are being generated, our advanced models identify more supporting web pages, allowing us to display a wider and more diverse set of helpful links associated with the response than with a classic web search.” — Google Search Central, AI features and your website. Jump to quote

Microsoft Bing — surfacing the retrieval layer

• “We are happy to introduce AI Performance in Bing Webmaster ToolsMicrosoft's free portal for monitoring and improving how a site appears in Bing search — the peer to Google Search Console, plus IndexNow instant indexing, richer backlink data, and keyword volumes. Because Bing's index also feeds Microsoft Copilot, it doubles as a window into AI-search visibility., a new set of insights that shows how publisher content appears across Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations.” — Bing Webmaster Blog, February 2026. Read the announcement

AI-citation readiness checklist

A pass to confirm you can be cited, then that you’re set up to see it:

Can you be retrieved and cited at all?

• Important pages are indexedStoring a crawled page in the search index so it can appear in results. Crawled is not the same as indexed — Google selects what to keep, and indexing isn't guaranteed. and snippet-eligible (the AI-citation floor).
• You are not blocking the AI botsAI crawlers are bots from AI companies that fetch web pages to train language models, build AI-search indexes, or answer live user questions. They come in three categories, each with its own user-agent tokens and its own robots.txt controls. you want to reach you — Googlebot-Extended, OAI-SearchBot, PerplexityBot, and the -User inference botsA crawler — also called a spider or bot — is an automated program that fetches web pages, extracts their links, and queues new URLs to visit. Search engines use crawlers to discover and download content for their index. — in robots.txt.
• Live-inference bots (ChatGPT-User, Perplexity-User, Claude-User) get clean 200s, not errors — every error is a missed real-time citation.
• Content isn’t gated behind heavy client-side JS that AI fetchers won’t render.

Are your pages the obvious source for a claim?

• Key answers are self-contained and appear early on the page.
• Headings mirror the specific questions (and sub-questions) people ask.
• The page offers non-commodity substance — a unique take, data, or first-hand experience — not restated common knowledge.

Are you building the off-site signal that actually drives citation?

• You’re tracking branded web mentions, not just backlinks.
• You have a presence on the third-party sources AI pulls from (industry publishers, Reddit, YouTube, review platforms).

Can you measure all three states?

• Log analysis in place for retrieved.
• Brand monitoring in place for mentioned.
• Citation tracking (Brand Radar cited reports, Bing WMT AI Performance) for cited.
• You’re checking each prompt multiple times, knowing AI outputs are volatile.

The mental models

1. Retrieved → Mentioned → Cited (→ Clicked). Four states, each strictly harder to reach than the last, each needing a different tool to observe. When AI visibilityLLM visibility (or AI visibility) is the aggregate measure of how often and how prominently a brand or page shows up in AI-generated answers — across AI Overviews, ChatGPT, Perplexity, Copilot, and Gemini. It's the AI-search analog of organic visibility, but it's driven by different signals. feels “broken,” locate which transition is failing — access, attribution, reputation, or click — because the fix for each is unrelated to the others.

2. Citation eligibility is classic SEO; citation selection isn’t. To be eligible, you need the old gates: indexedStoring a crawled page in the search index so it can appear in results. Crawled is not the same as indexed — Google selects what to keep, and indexing isn't guaranteed., snippet-eligible, crawlable. To be selected, rank position appears to matter less over time (a reported, not fully re-verified, 76% → 38% top-10 share) and being-talked-about matters more. Hold both ideas at once.

3. Mentions beat links for AI. Branded web mentions predict AI visibility (~0.66) better than Domain Rating (~0.33). The proxy that best matches what AI rewards is “is this brand discussed on authoritative pages,” not “is this URL linked.” Optimize the conversation, not just the link graph.

4. Your site is a minority of your own citations. ~90% of citations about a brand come from third-party sites; your own domain is 5–10%. You can’t on-site your way to AI visibility — most of the surface area is other people’s pages talking about you.

5. Cited is not a traffic channel (yet). Most AI sessions end without a click. Treat citation as brand presence in the answer layer — and make sure you’re named alongside the link, or the association is lost.

6. It’s probabilistic, so measure in aggregate. The same prompt yields different sources on different runs. One check is noise; run prompts repeatedly and trend the aggregate.

AI citations — cheat sheet

The three (four) states

BotsA crawler — also called a spider or bot — is an automated program that fetches web pages, extracts their links, and queues new URLs to visit. Search engines use crawlers to discover and download content for their index.: training/indexStoring a crawled page in the search index so it can appear in results. Crawled is not the same as indexed — Google selects what to keep, and indexing isn't guaranteed. vs. live-inference

A -User hit ≈ “the engine fetched this page to potentially cite in a live answer.”

Citation ≠ ranking ≠ backlinks — fast facts

• Top-10 organic share of AI OverviewAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index. citations: reportedly ~76% (mid-2025) → ~38% (early 2026) (Ahrefs; exact methodology comparison unverified — treat the direction as the signal).
• Branded web mentions ≈ 0.66 correlation with AI visibilityLLM visibility (or AI visibility) is the aggregate measure of how often and how prominently a brand or page shows up in AI-generated answers — across AI Overviews, ChatGPT, Perplexity, Copilot, and Gemini. It's the AI-search analog of organic visibility, but it's driven by different signals. vs. ~0.33 Domain Rating.
• ~90% of citations about a brand are on third-party sites (own site: 5–10%).
• No separate AI index — indexed + snippet-eligible is the floor for Google.
• Platform citation behavior: Perplexity cites almost everything it retrieves; ChatGPT cites a small fraction and often answers from training data.
• Cited ≠ click: in Pew’s March 2025 U.S. browsing study, users clicked a cited source in 1% of Google visits with an AI summary.

Playbook: turning “not cited” into “cited”

A staged plan, run in order — earlier stages gate later ones.

Stage 1 — Guarantee access (this week). Pull server logsLog file analysis is reading a web server's raw access logs to see exactly which URLs search engine crawlers actually requested, when, how often, and what status code they got. Unlike crawl tools or Search Console, logs are the unsampled, ground-truth record of what really happened., confirm AI botsAI crawlers are bots from AI companies that fetch web pages to train language models, build AI-search indexes, or answer live user questions. They come in three categories, each with its own user-agent tokens and its own robots.txt controls. are reaching you, and confirm they get 200s. Un-block any AI fetchers you actually want (default: don’t block them). Fix errors served to -User inference botsA crawler — also called a spider or bot — is an automated program that fetches web pages, extracts their links, and queues new URLs to visit. Search engines use crawlers to discover and download content for their index. first — those are live, in-conversation misses.

Stage 2 — Make the page the obvious source (this month). For each priority topic, ensure the specific answer is self-contained, early, and under a heading that mirrors the real question. Add the non-commodity layer — data, a first-hand result, an opinion — that a generic restatement can’t match. This is what moves you from retrieved to cited.

Stage 3 — Build the off-site conversation (ongoing). Because ~90% of your citations live on third-party sites, invest where AI actually pulls from: get mentioned in industry publications, seed genuine presence on Reddit and YouTube, earn review-platform coverage. The goal is branded web mentions on authoritative pages, not just links. This is the highest-leverage stage and the slowest.

Stage 4 — Reclaim hallucinated demand. AI engines sometimes invent plausible-but-nonexistent URLs on your domain. Find them (logs, 404 reports) and 301 them to the right live page so the traffic — and the implied citation intent — isn’t wasted.

Stage 5 — Measure the whole funnel, in aggregate. Stand up the three-layer stack (logs → brand monitoring → citation tracking), run priority prompts repeatedly, and trend the aggregate rather than reacting to any single volatile response.

What not to do

Assuming rank #1 = cited. Increasingly false — Ahrefs research puts the top-10 organic share of AI OverviewAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index. citations at roughly ~38%, down from a reported ~76% (exact methodology across the two studies still unverified, but the direction is consistent with query fan-outQuery fan-out is the technique where an AI search system breaks a single user question into multiple related sub-queries, runs those searches concurrently, and synthesizes the retrieved results into one answer. Google confirms AI Overviews and AI Mode 'may use a query fan-out technique' issuing multiple related searches across subtopics.). A #1 page has no guarantee; a page outside the top 100 can still be cited (see YouTube).

Treating backlinks as the AI-citation lever. Branded web mentions outpredict Domain Rating for AI visibilityLLM visibility (or AI visibility) is the aggregate measure of how often and how prominently a brand or page shows up in AI-generated answers — across AI Overviews, ChatGPT, Perplexity, Copilot, and Gemini. It's the AI-search analog of organic visibility, but it's driven by different signals.. Chasing links while ignoring whether you’re talked about optimizes the wrong signal.

Only on-site optimization. Your own domain is ~5–10% of the citations about your brand. If your entire AI strategy is your own pages, you’re ignoring ~90% of the surface area.

Reading citation-only tools as your “full AI presence.” They see the cited state only. Retrieval and mentions stay invisible — you can be shaping answers heavily with almost no visible citations.

Trusting a citation as automatic proof of support. A citation is attribution, not verification. Read the cited page against the specific claim before you treat it as backed up.

Confusing GPTBot with ChatGPT-User. A training-crawl hit tells you little about live citation; the -User inference botA crawler — also called a spider or bot — is an automated program that fetches web pages, extracts their links, and queues new URLs to visit. Search engines use crawlers to discover and download content for their index. is the signal that matters. Blocking the wrong one (or blocking inference bots outright) can quietly remove you from answers.

Blocking AI botsAI crawlers are bots from AI companies that fetch web pages to train language models, build AI-search indexes, or answer live user questions. They come in three categories, each with its own user-agent tokens and its own robots.txt controls. by reflex. Block the inference fetchers or Googlebot-Extended and you remove yourself from retrieval, mention, and citation at once. For most publishers that’s self-harm.

Judging visibility off a single prompt run. AI outputs are volatile — the same prompt yields different brand lists and sources across runs. One check is noise.

Chopping content into tiny “AI-friendly” chunks. Google explicitly says there’s no requirement to break content into tiny pieces; clear structure on a normal page is enough. (Same for llms.txtllms.txt is a proposed (not adopted) Markdown file at /llms.txt that gives AI systems a curated map of a site's most important pages. Proposed by Jeremy Howard in 2024, it's read mostly by coding agents like Claude Code — not search crawlers — and Google ignores it. and “special” schema — Google doesn’t use them for this.)

See who’s retrieving you: grep AI bots in your logs

Retrieval is the one state you can only observe in server logsLog file analysis is reading a web server's raw access logs to see exactly which URLs search engine crawlers actually requested, when, how often, and what status code they got. Unlike crawl tools or Search Console, logs are the unsampled, ground-truth record of what really happened.. These pull the AI crawlerA crawler — also called a spider or bot — is an automated program that fetches web pages, extracts their links, and queues new URLs to visit. Search engines use crawlers to discover and download content for their index. user-agentsA user agent is the HTTP request header a client (browser, crawler, or bot) sends to identify itself. For crawlers, a short user-agent token — a substring of that string — is what robots.txt rules actually target. out of a standard access log.

macOS / Linux (shell)

# All major AI bots, with a count per user-agent
grep -iE "GPTBot|OAI-SearchBot|ChatGPT-User|PerplexityBot|Perplexity-User|ClaudeBot|Claude-User|Googlebot-Extended" access.log \
  | grep -oiE "GPTBot|OAI-SearchBot|ChatGPT-User|PerplexityBot|Perplexity-User|ClaudeBot|Claude-User|Googlebot-Extended" \
  | sort | tr 'A-Z' 'a-z' | uniq -c | sort -rn

# The signal that matters most: live-inference (-User) bots that hit an ERROR
grep -iE "ChatGPT-User|Perplexity-User|Claude-User" access.log \
  | awk '$9 ~ /^(4|5)/ {print $9, $7}' | sort | uniq -c | sort -rn

Regex for a log platform (Splunk / BigQuery / Loki)

(?i)\b(gptbot|oai-searchbot|chatgpt-user|perplexitybot|perplexity-user|claudebot|claude-user|googlebot-extended)\b

Split the alternation into a “training/indexStoring a crawled page in the search index so it can appear in results. Crawled is not the same as indexed — Google selects what to keep, and indexing isn't guaranteed.” group vs. a “-User inference” group to separate crawl noise from real-time, in-conversation fetches.

Extract the source links from an AI answer (browser)

Paste into Chrome DevTools Console on an AI answer page to list the cited source domains — a quick way to check who got the citation slots for a query.

// Console snippet: unique cited domains among outbound links on the page
[...new Set([...document.querySelectorAll('a[href^="http"]')]
  .map(a => { try { return new URL(a.href).hostname.replace(/^www\./,''); }
             catch { return null; } })
  .filter(Boolean))]
  .sort()
  .forEach(d => console.log(d));

Bookmarklet: highlight your own domain in a set of citations

Drop this in a bookmark and click it on an AI answer to outline any link pointing at your site — fast visual check for “was I cited here?”

javascript:(()=>{const dom='example.com';document.querySelectorAll('a[href*="'+dom+'"]').forEach(a=>{a.style.outline='3px solid #85BB65';a.style.background='rgba(133,187,101,.15)';});})();

Swap example.com for your domain. (Logs still tell you about retrieval; this only inspects visible citations in the rendered answer.)

Tools for seeing each state

Retrieved (logs)

• Server log file analysers — Screaming Frog Log File Analyser (has an AI-botA crawler — also called a spider or bot — is an automated program that fetches web pages, extracts their links, and queues new URLs to visit. Search engines use crawlers to discover and download content for their index. monitoring workflow), Botify, and Ahrefs Bot Analytics break traffic down by individual AI botAI crawlers are bots from AI companies that fetch web pages to train language models, build AI-search indexes, or answer live user questions. They come in three categories, each with its own user-agent tokens and its own robots.txt controls. and by category. This is the only place retrieval is visible.
• CDN / edge analytics (e.g. Cloudflare) — bot breakdowns if you don’t have raw logs.

Mentioned (response analysis)

• Ahrefs Brand Radar — monitors AI responses across engines and surfaces brand mentions as string matches, including a “foundA 302 (\"Found\") is a temporary redirect: it forwards users to a new URL while telling search engines the original URL should stay in the index. It's a weak canonicalization signal, not the zero-equity dead end of SEO folklore., but not cited” view (mentioned without a link to you).

Cited (citation tracking)

• Ahrefs Brand Radar — Cited Domains / Cited Pages — which third-party sites get cited when you’re discussed, and which of your own pages get cited most.
• Bing Webmaster ToolsMicrosoft's free portal for monitoring and improving how a site appears in Bing search — the peer to Google Search Console, plus IndexNow instant indexing, richer backlink data, and keyword volumes. Because Bing's index also feeds Microsoft Copilot, it doubles as a window into AI-search visibility. — AI Performance — official Copilot citation counts, average cited pages, and groundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it. queries; the most reliable source for Copilot.
• Google Search ConsoleA free Google service that reports how a site performs in Google Search and surfaces problems with how Google crawls, indexes, and serves it. It's first-party data straight from Google — but you don't need it to appear in results. — Generative AI performance reportThe Google Search Console report that shows how your site actually performed in Google Search, built from real impressions and clicks. It reports four metrics — clicks, impressions, average CTR, and average position — and keeps the most recent 16 months of data. — impressions-only visibility for AI features (no clicks/CTR/query dimension yet).

Reality check

• No single tool sees all three states — you need the stack, not one dashboard.

Monthly AI citation review

1. Freeze the prompt list, platform/model names, geography, account state, and run date.
2. Run the same prompts in fresh sessions and save complete answers with visible source links, not screenshots of a cropped citation tray.
3. Normalize every cited URL to its final canonical destination and record the source page, prompt, platform, position, and whether the brand was mentioned in the answer.
4. Separate citation, unlinked mention, retrieval evidence from logs, and referral visit.
5. Verify that cited pages support the adjacent claims; flag mismatches and fabricated URLs for manual review.
6. Compare with the prior run by prompt cohort and platform. Do not combine changing prompts into an apparent visibility trend.
7. Assign actions to content, technical, digital PR, or measurement owners and archive the raw evidence with the report.

Validate citation measurement

AI citation metrics

Resources worth your time

My related writing & research

• 76% of AI Overview Citations Pull From the Top 10 — my original citation-vs-ranking study the update builds on.
• Insights From 55.8M AI Overviews Across 590M Searches — my large-scale look at where AI OverviewsAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index. show and which domains get cited.

My speaking

• GEO? AEO? LLMO? What’s With All This AI SEO Stuff? (Speaker Deck) — my Ahrefs Evolve 2025 talk, source of the “optimize how the internet talks about you” framing and the Zapier-mentions-Ahrefs citation-leverage example. Video.
• Cutting-Edge AEO Strategies with Patrick Stox (Marketing Speak, ep. 539) — a longer conversation on the retrieved/mentioned/cited reality.

From around the industry

• What We Actually Know About Optimizing for LLM Search — an Ahrefs overview of the 75,000-brand study by Louise Linehan and Xibeijia Guan, including the branded-mention and Domain Rating correlations.
• Update: 38% of AI Overview Citations Pull From The Top 10 — Xibeijia Guan’s March 2026 study reporting the top-10 citation share falling from ~76% to ~38% (the exact methodology comparison against the earlier study hasn’t been independently reverified here), and YouTube as the most-cited domain.
• Generative Engine Optimization — Ahrefs’ strategy guide to getting cited.
• Evaluating Verifiability in Generative Search Engines — Liu, Zhang & Liang (Stanford, EMNLP 2023): only ~51.5% of generated sentences are fully supported by their citations, and ~74.5% of citations actually support their sentence.
• Introducing AI Performance in Bing Webmaster Tools — Microsoft’s official citation + groundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it.-queries report.
• The AI Citations Report 2026 — OtterlyAI (1M+ data points): the “99.6% of AI influence is invisible” groundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it. stat and that ~73% of sites have crawlability issuesCrawlability is how well search engine crawlers can discover, access, and fetch a site's pages. A crawlability issue is any technical condition — blocked access, broken links, server failures, or bloated URL inventory — that stops pages from reaching the index. blocking AI access.
• AI-are-highly-inconsistent-when-recommending-brands — SparkToro: less than a 1-in-100 chance of identical brand lists across 100 runs of the same prompt.
• How each platform cites sources differently — Discovered Labs: ChatGPT cites only ~15% of what it retrieves; ~82.9% of B2B citations come from third-party sources.
• Monitor AI bots in the Log File Analyser — Screaming Frog on the -User vs. training-botA crawler — also called a spider or bot — is an automated program that fetches web pages, extracts their links, and queues new URLs to visit. Search engines use crawlers to discover and download content for their index. distinction, and why every error to an inference bot is a missed citation.
• How AI Search Engines Work — Ryan Law (Ahrefs) on RAGRAG is the retrieve-then-generate pattern behind AI search: the system retrieves relevant passages from an external index at query time, injects them into the model's context, and generates an answer grounded in those sources — without changing the model's weights. as the grounding mechanism behind citations.

Stats worth citing

• ~76% → ~38% (unverified methodology comparison) — reported share of AI Overview citations coming from the organic top 10, mid-2025 vs. early 2026; ranking and AI citation appear to be decoupling as query fan-outQuery fan-out is the technique where an AI search system breaks a single user question into multiple related sub-queries, runs those searches concurrently, and synthesizes the retrieved results into one answer. Google confirms AI Overviews and AI Mode 'may use a query fan-out technique' issuing multiple related searches across subtopics. matures, but the exact denominators behind this specific comparison haven’t been independently reconfirmed. Source
• ρ ≈ 0.664 vs. ~0.326 — branded web mentions vs. Domain Rating as predictors of AI OverviewAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index. inclusion in Ahrefs’ 75,000-brand study by Louise Linehan and Xibeijia Guan; mentions win. Source
• ~5–10% — the share of sources AI platforms reference about a brand that live on the brand’s own site; ~90% are third-party (McKinsey, relayed). Context
• ~15% — the fraction of retrieved pages ChatGPT actually cites; being retrieved is necessary but far from sufficient (Discovered Labs, relayed). Source
• 51.5% / 74.5% — of AI-generated sentences fully supported by citations, and of citations that actually support their sentence, respectively (Stanford, EMNLP 2023). Source
• ~1% — of users who click a cited source within an AI summary; most AI sessions end without a site visit (Pew, relayed). Cited rarely means clicked. Context

Test yourself: AI Citations

Five quick questions on how AI answer enginesAnswer Engine Optimization (AEO) is the practice of structuring content so engines deliver it as a direct answer — featured snippets, voice assistants, and AI search — rather than just a ranked link. Coined for voice search in 2018 and revived for the LLM era. Google's position is that it's still SEO. choose what to cite. Pick an answer for each, then check.

AI Citations

An AI citation is the visible source link an AI answer engine shows next to its generated text — the clickable reference that credits the web page it used. A citation's presence is a separate thing from whether the cited page actually supports the statement, and from being retrieved (read behind the scenes) or merely mentioned (named without a link); citation is driven more by brand mentions and being retrievable than by traditional ranking.

Related: Retrieval-Augmented Generation (RAG), Grounding

AI Citations

An AI citation is the visible, usually clickable source link an AI answer engine — Google’s AI OverviewsAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index. and AI Mode, ChatGPT Search, Perplexity, Microsoft Copilot — shows to credit a web page it used to build its response. It’s the AI-search equivalent of a search result link, but it’s chosen at generation time from the sources the system retrieved, not ranked into a list of ten blue links. A citation being shown is not proof the cited page actually supports the claim next to it — that’s a separate, claim-level check.

Citation sits at the end of a chain of three distinct states, and conflating them is where most measurement goes wrong. Retrieved means a request for your page shows up in server logsLog file analysis is reading a web server's raw access logs to see exactly which URLs search engine crawlers actually requested, when, how often, and what status code they got. Unlike crawl tools or Search Console, logs are the unsampled, ground-truth record of what really happened. — real evidence of a fetch, but not proof the content reached the model’s context or was used. Mentioned means your brand name appears in the generated text with no link. Cited means your URL is shown as an attributed source. Content can be retrieved without being mentioned, and mentioned without being cited — each state needs a different tool to see it.

Two things make AI citations behave unlike classic ranking. First, there’s no separate “AI indexStoring a crawled page in the search index so it can appear in results. Crawled is not the same as indexed — Google selects what to keep, and indexing isn't guaranteed.”: Google grounds AI OverviewsAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index. in its core Search index via Retrieval-Augmented GenerationRAG is the retrieve-then-generate pattern behind AI search: the system retrieves relevant passages from an external index at query time, injects them into the model's context, and generates an answer grounded in those sources — without changing the model's weights., so a page must already be indexed and snippet-eligible to be cited. Second, being cited correlates weakly with ranking #1 — Ahrefs research points to the share of AI Overview citations coming from the organic top 10 falling from roughly 76% in mid-2025 to about 38% by early 2026 as query fan-outQuery fan-out is the technique where an AI search system breaks a single user question into multiple related sub-queries, runs those searches concurrently, and synthesizes the retrieved results into one answer. Google confirms AI Overviews and AI Mode 'may use a query fan-out technique' issuing multiple related searches across subtopics. matured, though the exact methodology behind that specific comparison hasn’t been independently reverified. In practice, branded web mentions on third-party sites predict AI visibilityLLM visibility (or AI visibility) is the aggregate measure of how often and how prominently a brand or page shows up in AI-generated answers — across AI Overviews, ChatGPT, Perplexity, Copilot, and Gemini. It's the AI-search analog of organic visibility, but it's driven by different signals. more strongly than Domain Rating does — which is why the work has shifted from optimizing your site to optimizing how the internet talks about you.

Related: Retrieval-Augmented Generation (RAG), Grounding

Changelog