We Rank #1 on Google But Don't Show Up in AI Answers

Six overlap statistics are circulating and they contradict each other. Here is what each one actually counts, and a 20 minute rule for telling whether you have a page problem or a brand problem.

Last updated: 30 August 2026

"My Google rankings and paid ads are solid, but we're completely invisible in ChatGPT," from r/AskMarketing. "Ran an AEO audit for a client that sits top three on Google for its main terms. In ChatGPT and Perplexity, it barely came up," from a Konabos co-founder on LinkedIn.

Same complaint, different rooms. Go looking for the explanation and you hit six numbers that flatly disagree. Overlap collapsed from 70% to under 20%. Or 76% to 38%. Or it was only ever 12%. Or 62%. Or it is rising. Every page picks one and moves on, and none tells you the six are not measuring the same thing.

One disclosure first. On 24 August 2026 I measured 40 B2B buyer questions across six AI engines, 240 measurement points. My own domain was cited zero times. Not the winner's seat.

Why doesn't ranking #1 get me cited in AI answers?

Google published the mechanism on 15 May 2026 and named it query fan-out: "A set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results to address the user's query." Its own example fans "how to fix a lawn that's full of weeds" into "best herbicides for lawns," "remove weeds without chemicals," and others, then merges the results. A page ranking sixth across several of those variants can beat a page ranking first for only the literal question.

What gets matched is smaller than a page. Retrieval-augmented generation conditions the answer on a dense index of passages and "can use different passages per token" (Lewis et al., NeurIPS 2020), scored as embeddings and re-ranked. A November 2025 preprint from Dell Technologies' Services AI Research Group argues the two jobs should be split, setting the matching unit at "small, semantically dense segments (roughly 100-200 tokens)" and the assembly unit at "larger segments (roughly 600-1000 tokens)," and notes that most current pipelines collapse both into a single representation.

So the thing competing is not your page. It is a passage inside it, against sub-queries your buyer never typed. In ChatGPT, Google rank barely predicts where a brand lands in the answer: a correlation of 0.034 with browsing on and 0.022 with browsing off (Chatoptic, 15 brands, 1,000 queries). The coefficient type is not published, and the study covers ChatGPT alone. That is not a weak relationship. That is no relationship.

Which overlap statistic is actually right?

Read down the third column and the fight disappears. Ahrefs' 76% and 12% are one research house asking the same question on two different surfaces: AI Overviews follow the SERP, standalone assistants do not. BrightEdge's flat 17% and rising 54.5% are one surface measured at two depths: AI Overviews cite more ranked content now, but the growth is all in the deeper band, positions 11 to 100, while page-one overlap has not moved.

Two things almost every competing page omits. Ahrefs did flag that its own parsing improved between the two studies, "Since that last study we've also improved our parsing methodology in Ahrefs, so that we can see even more of the citations that appear in AI Overviews," but it raises this as a reason to redo the study rather than as a caveat on the comparison. The caveat is Search Engine Journal's, in its write-up, which reads the parsing change as a limit on comparing the two datasets. I had it as an Ahrefs quotation in an earlier draft, and it is not one. The direction is real; the size of the fall is not cleanly measured. And Search Engine Land, owned by Semrush, runs both the "62%, near-zero correlation" story and BrightEdge's convergence framing without reconciling them.

The most defensible sentence available: ranking predicts citation strongly on Google AI Overviews and weakly to not at all on standalone assistants. There is no single overlap number.

A seventh number, with the method attached

On 24 August 2026 I logged every cited source across all six engines for 40 B2B buyer questions, alongside the organic top three per query. Organic results came back for 34 of the 40, giving 102 top-three domain slots. Of those, 63 also appeared as a cited source from at least one engine on the same query: 61.8%, within a point of Chatoptic's 62%, and for the same reason. It is a generous unit, domain level and any engine.

Narrow the unit and the same dataset answers differently every time.

One window, one dataset, a range from 4.4% to 61.8%. The number is decided by the unit and surface you picked before counting.

Limits: one run per engine per query, domain level not URL level, 34 queries. By my own rule below, one run cannot call an absence real, so treat it as directional. On ChatGPT it lines up with published work on rank order, mine is last of six and Ahrefs has it lowest too. On Gemini it cannot agree or disagree: my 29.6% is ranking-to-citation at domain level across the organic top three, Ahrefs' 8% is citation-to-ranking at URL level against the top ten. Two numbers about the same engine, 21 points apart, that are not measuring the same thing. That is the whole argument in one row.

Two widely repeated claims did not survive checking, so they are absent above. The 5WPR figure has no traceable methodology. And "structured data means 3.2x more citations," attributed to a Princeton paper, is not in that paper, which never tested schema.

A third is not unsourced, it is misfiled, and I had it in the wrong bucket. "Pages above 20,000 characters average 10.18 citations" traces to Kevin Indig's Growth Memo study of 23 March 2026, roughly 98,000 citation rows drawn from about 1.2M ChatGPT responses. It is routinely quoted against Ahrefs' 0.04 word-count correlation as though one refutes the other. They cannot. Indig measured ChatGPT, counted characters, and reported mean citations per page. Ahrefs measured Google AI Overviews, counted words, and reported a correlation among pages already cited. Different engine, different unit, different question, which is the same failure as the six numbers above.

The study that would settle this does not exist: nobody has compared a cited #1-ranked page against an uncited one, holding query and position constant. Every named before-and-after case I found is vendor-published, not independently audited.

Is this a page problem or a brand problem?

1. Is the page indexed and eligible to show in Search with a snippet? Google makes this a hard prerequisite, along with the site being enabled for generative AI features in Search Console. If no, stop. Page problem, and a cheap one.
2. Does it rank top 10 for the target question and for three to five fan-out variants of it? Ranking for the head term but not the variants is still a page problem: you lack passage-level coverage of the adjacent intents the engine retrieves against.
3. Run the buyer question 10 times in ChatGPT and Perplexity. Are competitors named while you are absent, despite your ranking? Then it is a brand problem. In ChatGPT, Google position correlates around 0.03 with where a brand lands in the answer, and about 80% of standalone-assistant citations are not from pages ranking for that query, so more ranking will not fix it.
4. Is your brand named, but the citation points at G2 or Reddit rather than your site? Corroboration exists and your page is losing the passage. Fix extractability, keep the third-party presence.

Fail 1 or 2 and it is a page problem, fast and cheap. Pass those but fail 3 and it is a brand problem, slow and expensive. The rule is synthesised from the evidence, not itself tested, so treat it as triage.

Do the AI engines cite differently?

Ahrefs found AI Mode reproduced every entity the AI Overview named, and added more, in 61% of response pairs, which it summarises as roughly a 61% chance your brand carries over. But that is entity level. At URL level the same study puts the two surfaces at 13.7% shared citations. Same pair of surfaces, same dataset, 47 percentage points apart, because one counts brands and the other counts pages. On my own 40 queries the split was starker: on this exact question, AI Mode returned 35 sources and AI Overview none.

The B2B source pool is not where most teams assume. Foundation Inc, across 50 brands and seven verticals, found Reddit is 20.8% of top-50 external citation domains and the largest external source in six of seven; review sites like G2 are 4%. YouTube is the most-cited AI Overview domain, but read the denominator: 21.1% is its share of the citations going to the top 50 sources, while in the 863,000-keyword dataset behind the 38% row above it is 5.6% of all AI Overview URLs cited. Same domain, same research house, two denominators, a 4x gap. In that second dataset, 18.2% of AI Overview citations from outside Google's top 100 are YouTube URLs. One caveat changes the plan: ChatGPT's Reddit citations collapsed in August 2026. Promptwatch measured Reddit's share of ChatGPT Search citations falling from 3.83% in late July to 0.52% in mid-August, and attributes it to a retrieval change rather than a Reddit block. Across my own 551 ChatGPT and 396 Copilot sources on 24 August, zero were Reddit or Quora. But note what each number counts: Ahrefs still had reddit.com as ChatGPT's most-cited domain in July 2026, at 16.7% of the citations going to the top 50 sources. Below one per cent and still number one are both true, because they count different denominators over different months. Reddit buys the Google surfaces and Perplexity; on ChatGPT, plan for owned pages.

What do I fix first, and who on my team owns it?

• Engineer. Indexation, Search snippet eligibility, Search generative AI control set to include, under Search Console settings. Prerequisite, not optimisation. Include is the default, but child properties inherit the parent, so a domain-level exclude propagates silently. A thirty-second check, not a task. Then JavaScript rendering so content sits in visible HTML, plus crawlability and Bing Webmaster Tools for Copilot and ChatGPT.
• Writer. Answer first under every heading. For chunk-retrieval assistants — ChatGPT, Perplexity — a 40 to 70 word answer-first passage that stands alone is the habit worth building. Google's own guide says the opposite pressure doesn't exist on its surfaces: you "don't need to write in a specific way just for generative AI search." Google says content people find unique, compelling and useful will likely influence a site's presence in generative AI search more than any other suggestion in its guide, and names a unique point of view and non-commodity content as two of its attributes.
• PR and demand generation. G2, Capterra and TrustRadius, a real Reddit presence, Wikidata and Wikipedia entity identity, consistent brand naming.
• Executive. Fast page fixes first, slow corroboration in parallel, then decide what to stop.

What to stop, per Google's own guide: llms.txt and other special markup, because "Google Search itself doesn't use them." Chunking content into tiny pieces: "there's no requirement to break your content into tiny pieces for AI to better understand it." Obsessing over word counts: "there's no ideal page length." Rewriting content just for AI. Seeking inauthentic mentions. Over-focusing on structured data, which "isn't required."

On schema the correlation is real and the causation is not. Ahrefs found cited pages roughly 3x more likely to carry JSON-LD across 6M URLs, then ran the controlled version: 1,885 pages adding schema against about 4,000 matched controls. It "produced no major uplift in citations on any platform," on pages AI was already citing heavily; Ahrefs notes schema may still matter for pages AI has never surfaced. Keep schema for rich results, not as a citation lever.

What it costs. Plenty of agencies publish price tables; none publishes a survey with a sample behind it, so treat all of this as rate cards rather than market data. Self-serve tools $29-$489 a month, mid-market retainers $2,000-$8,000, entry work $1,000-$2,500, enterprise $10,000 and up, freelancers $75-$150 an hour, a full-time specialist $70,000-$110,000 a year in the US. The commonest classic SEO retainer is $500-$1,000 a month.

How long each layer takes to show effect, not how long it lasts. One row is documented, the re-crawl, per Google Search Central. The rest are practitioner estimates from agency guides with no published methodology, so treat them as planning ranges rather than measurements:

Once I'm cited, does it stay?

Those facts are usually reported separately and never joined. Between 40% and 60% of cited domains rotate month to month, 70% to 90% over six months. Semrush found ChatGPT citing Reddit in close to 60% of prompt responses in early August 2025, collapsing to around 10% by mid-September. Note the unit, because it is the whole lesson: that is the share of responses containing a Reddit citation, not Reddit's share of all citations. Profound measured the same event the other way round and got about 7% falling to about 1%. One event, two honest numbers an order of magnitude apart. Only 30% of brands visible in one answer remain in the next.

Which has a consequence nobody states: one absent answer proves nothing. Two repeat ChatGPT answers to the same prompt share only 21.2% of their cited domains, and the one paper to study run counts directly, Schulte, Bleeker and Kaufmann (arXiv:2604.07585, April 2026), concludes that answers vary enough across runs that one-off observations are unreliable and repeated measurement is required. It does not name a magic number, and neither will I. Run 10 per prompt per engine and call yourself absent only at 0 of 10. And log three units separately, never collapsed, because collapsing them is exactly what produced the six contradictory statistics above: a mention is your name in the answer text, a citation is your URL as a linked source, a recommendation is your brand on a shortlist.

One last thing about the ranking you already hold. Ahrefs put position-one CTR 34.5% lower on keywords carrying an AI Overview than on comparable keywords without one, which is a relative gap against a counterfactual, not percentage points. And the "58% at position one against 14% at position 10" line that usually follows is not about Google at all: it is ChatGPT's own fan-out search rank, 58.4% against 14.2%, from AirOps with Kevin Indig across 16,851 queries in April 2026. Your Google position is not what that figure measures.

Frequently asked questions

How do we get into Google AI Overviews?

Be indexed, eligible to show in Search with a snippet, and not excluded by the Search generative AI control in Search Console. Then rank for the query cluster, not just the head term. Google states AI features are "rooted in our core Search ranking and quality systems."

What content changes improve AI answer visibility?

Answer-first structure under each heading — for chunk-retrieval assistants like ChatGPT and Perplexity, a 40 to 70 word passage that makes sense lifted out of the page; Google itself says no special writing style is required for its AI surfaces. Structure decides this, not length: Ahrefs, analysing 560,346 Google AI Overviews, found no meaningful link between length and citation across 174,048 cited pages, a 0.04 correlation between word count and citation position, and 53.4% of those cited pages are under 1,000 words.

Does schema markup help with AI answers?

Not causally. Ahrefs' controlled test of 1,885 pages adding JSON-LD against matched controls found "no major uplift in citations on any platform," and Google states structured data "isn't required."

How do we track AI overview rankings?

Search Console's generative AI report for Google's surfaces, then 30 prompts at 10 runs each per engine, logging mentions, citations and recommendations separately. Report frequency per engine, not blended. Add a GA4 custom channel group, since AI traffic otherwise lands in Other or Direct.

Why are competitors cited in AI answers instead?

Usually stronger third-party corroboration and clearer entity consensus, not better rankings. In ChatGPT, Google position correlates around 0.03 with brand-mention order (Chatoptic, 15 brands, 1,000 queries), and 28.3% of ChatGPT's most-cited pages have zero Google organic visibility.

Is it a page problem or a brand problem?

Work the four steps in order. If the page is not indexed and snippet-eligible, or does not rank for the fan-out variants as well as the head term, it is a page problem and cheap to fix. If it passes both and competitors are still named across ten runs while you are absent, it is a brand problem, and more ranking will not fix it.

How many times should I run a prompt before believing it?

Ten per prompt per engine, and call yourself absent only at zero of ten. Two repeat ChatGPT answers to the same question share only 21.2% of cited domains, so a single absent reading is mostly noise rather than evidence.

---

Sources

1. Google Search Central, generative AI optimization guide, 15 May 2026 (updated 10 Jul 2026).
2. Ahrefs: 12% overlap study (11 Aug 2025); 76% AI Overview study (Jul 2025); 38% update, 2 Mar 2026 (covered by Search Engine Journal the same day); schema controlled test (2026); content length study (2026); AI Overviews vs AI Mode (Dec 2025); Brand Radar most-cited domains (Jul 2026).
3. BrightEdge: 16-month rank overlap (2025); AI Overviews at one year (2025); trigger rate (Feb 2026).
4. Chatoptic, Google and ChatGPT visibility study, 2025.
5. 5WPR, "GEO vs. SEO: The 2026 Venn Diagram," 4 May 2026, via PR Newswire.
6. Pew Research Center, clicks on links in Google AI summaries, 22 Jul 2025.
7. Lewis et al., NeurIPS 2020, arXiv:2005.11401; Karpukhin et al., EMNLP 2020, arXiv:2004.04906; Nogueira and Cho, 2019, arXiv:1901.04085; Nainwani and Baban, Nov 2025, arXiv:2511.04939.
8. SE Ranking, AI Mode URL volatility, collected 20 Jun 2025 (10,000 keywords, three same-day runs); Parse, AI citation volatility, 8 Jul 2026 (16,143 ChatGPT prompts, source of the 21.2% repeat-answer overlap); Profound, AI Search Volatility, 17 Jul 2025 (source of the 40-60% and 70-90% rotation figures); Scrunch and Stacker source-decay research, 26 Mar 2026 (source of the 4.5-week and 10-week half-lives); AirOps and Kevin Indig, State of AI Search, Dec 2025 (the 30% carry-over figure); Schulte, Bleeker and Kaufmann, arXiv:2604.07585, 8 Apr 2026 (run-count reliability).
9. Foundation Inc, Reddit's share of B2B AI citations, 2026. Growth Memo, Apr 2026.
10. Search Engine Land, articles 461891 and 473325 (owned by Semrush).
11. Exalt Growth, GoCodes case study, 20 Aug 2026 (vendor-published).
12. Digital Elevator AEO/GEO pricing guide, 2026, for the retainer bands; Fuel Online AI SEO pricing, 2026, for the freelance and salary figures; Ahrefs 2024 agency survey for the classic SEO retainer. WebFX's 2026 guide gives materially higher agency bands and does not support the ranges above.
13. Paraphrase Labs AI Visibility Baseline, 24 Aug 2026: 40 B2B buyer questions, six engines, 240 measurement points, 1,515 citations across 1,028 domains.