The AI visibility funnel: five stages, five instruments, and the fine print the headline dropped

TL;DR

• The figure from Ahrefs’ Q1 2026 AI Search Benchmark that will get repeated is this one: 28.3 percent of the 1,000 pages ChatGPT cites most have no organic keyword visibility at all. Two figures in the same section change what it means. The 71.7 percent that do rank sit on domains with a median Domain Rating of 90, and Ahrefs classes only 32.3 percent of the cited pages as content types a brand could realistically influence. Rank is not the gate. Authority mostly still is.
• Length is not a lever. 53.4 percent of the pages cited in Google’s AI Overviews run under 1,000 words, and across 174,048 pages stripped of boilerplate, word count and citation showed near-zero correlation.
• Other people’s lists beat your own pages. “Best X” blog lists made up 43.8 percent of ChatGPT’s source URLs on 750 top-of-funnel queries.
• Measure in five stages (retrieved, cited, named, still visible, converting) and measure each one per platform. A single blended AI-visibility score hides which stage you fail, and reporting a small channel on visit volume makes it look like a failure whatever it actually does.
• Scope, plainly: every number here is Ahrefs’ measurement of the English-language commercial web, weighted to software, agency and product verticals, published by a company that sells AI-visibility tracking. The structure travels to a Quebec business. The numbers do not.

A lot of AI-visibility reporting collapses into one number: a share-of-voice score, a visibility index, the percentage of prompts where the brand showed up. One number cannot tell you why it is low. A business whose pages the crawler never reaches and a business that gets quoted without being named produce the same bad score, and they need opposite work. Splitting the problem into stages is what turns a disappointing score into a work order, and it extends the measurement section of our GEO overview.

The five-stage framing below comes from an infographic by Maja Voje (GTM Strategist) built on Ahrefs’ AI Search Benchmark for Q1 2026. Ahrefs is a vendor that sells AI-visibility tracking alongside its SEO tools, so this is commercial research, and we treat its numbers that way. We took every figure back to the report itself before using it here. That check caught two figures that do not hold up, covered further down.

Five questions, in order

Each stage asks a different question and needs its own instrument. The order is the point: a page no AI crawler can fetch cannot be cited, so citation and mention work on that page buys nothing until stage one is fixed. A model can still name a business from what it absorbed in training, without fetching anything, but that is the memory layer, and it moves on a different clock.

Two of the five carry most of the value. Stage three’s instruction to count linked and unlinked mentions separately names a blind spot in the standard SEO toolkit. An unlinked mention never shows up in a backlink report or a referral report, so a business can be named in AI answers every week and see nothing in its usual dashboards. The benchmark measures this for a single brand, Ahrefs itself, over 31,000-plus AI mentions, and on AI Overviews, Gemini and Copilot most of those mentions carried no link. That is one brand, so the exact rate does not transfer. The direction does: if you only count links, you are missing most of the signal, which is why this stage has to be checked by hand or with a mention tracker, not read off a backlink tool.

Stage five’s “conversion per visit, never visit volume” is the reporting rule that keeps a small channel from being written off.

Do you need to rank in Google to be cited by ChatGPT?

Ahrefs took the 1,000 pages ChatGPT cites most and checked each one’s organic search visibility. 283 had no organic keywords at all. That is the 28.3 percent, and it is a real finding: ranking in Google is not a precondition for being cited by ChatGPT. Google rankings and AI citations are different scoreboards, which is why “we rank first on Google” does not answer the question of whether an AI names you. It is also the empirical form of a split we drew in the post on how AI chooses what to cite: being fetchable is the documented gate, and ranking is not part of it.

Now read the rest of that section of the report. The 717 pages that do rank are not small sites that got lucky. They rank for a median of 279 keywords each and sit on domains with a median Domain Rating of 90, which Ahrefs describes as among the most authoritative on the internet. And when Ahrefs sorted the cited pages by content type, only 32.3 percent were types a brand could realistically influence: educational content 19.4 percent, reviews 5.8, news and media 5.2, blog posts and articles 1.9. The rest are page types Ahrefs did not count as open to influence at all. (The report leaves it unclear whether that share is of all 1,000 pages or only of the 717 that rank; the point holds either way.)

So the honest reading has two halves. “You do not need to rank to get cited” is true. “You do not need authority to get cited” is false, and it is the version that tends to get sold. For a small business with a modest domain, the 28.3 percent is not an opening on its own. What it does justify is measuring citation directly instead of inferring it from rankings, because the two can diverge.

Does word count affect AI Overview citations?

In Ahrefs’ study of 174,000-odd pages cited in Google’s AI Overviews, 53.4 percent were under 1,000 words, 30.6 percent ran 1,000 to 2,000, and only 16 percent went past 2,000. The average was 1,282 words. Ahrefs stripped the HTML and boilerplate from all 174,048 pages before counting and found the correlation between word count and being cited near zero, weak enough that it described the effect as generally not worth considering.

That cuts against the word-count targets that still show up in content briefs. If the AI argument for a content plan is “longer, more comprehensive pages”, this is a measurement on Google’s AI surface that word count did not predict citation. It does not say short pages win either: near zero runs in both directions. The variable left standing is what the page says, which is the same conclusion the paraphrase test in our post on citation concentration reached from a completely different method. Two unrelated measurements point the same way: budget for substance, not for length or rewording.

Get onto other people’s lists

In Ahrefs’ Q1 2026 benchmark, “Best X” blog lists made up 43.8 percent of the URLs ChatGPT drew on across 750 top-of-funnel queries. That is an average across three verticals: 51.4 percent in software, 40.6 in agencies, 39.5 in products. Local services were not among them, so read the number as a strong pattern in those markets rather than a forecast for yours.

The direction is still the useful part. Being listed on the roundups and comparisons a buyer already reads in your category outweighs publishing one more page about yourself. The same report found the strongest single correlate of AI brand visibility, across 75,000 brands, was mentions on YouTube (a correlation of about 0.737), while backlink counts and URL Rating correlated only weakly. That is a correlation, not a mechanism. Read together, both point at where the conversation about you happens: on other people’s pages and channels, not only on your own site. That matches what we found on the memory layer, where the only lever was earned third-party coverage built up over years. Some of those lists do sell their slots. A listing bought in order to be cited is a placement, not a recommendation, and it is not the work this section is about.

Measure per platform, and report conversion, not visits

Across 730,000 pairs of responses to the same queries, Ahrefs found that Google’s AI Mode and AI Overviews cited almost entirely different sources (13.7 percent URL overlap) while reaching the same conclusion 86 percent of the time. Two Google surfaces built on the same index, and they barely agree on who to cite. A blended score averages that away. Keep each platform as its own row: if you are cited in one and absent from another, that is a finding, not noise to smooth over. The overlap figure also bears on stage four, with a limit: it compares two surfaces at one moment, not one surface over time. Even so, when two Google surfaces agree on the answer while citing different pages, the realistic target is to be part of the consensus a category forms, not to own one citation slot.

On the conversion stage, scale matters before rate. Across 76,000 sites running Ahrefs Web Analytics, Google sent 190 times more visitors than ChatGPT. Google accounted for nearly 40 percent of total traffic and ChatGPT for 0.21 percent, even though, counting only the prompts Ahrefs classes as search, ChatGPT handles roughly 12 percent of Google’s query volume. Ahrefs’ Patrick Stox attributes the gap mostly to a much lower click-through, around 96 percent lower by his estimate, because the assistant keeps the user in the conversation.

Reported on visit volume, then, AI search will look like a rounding error for most small businesses for some time. Report it on what each visit does: AI referral sessions by landing page, and the conversion rate per visit. That is a reporting rule, not a prediction that the rate will be high. Whether AI visitors convert better on your site is something to measure on your site, not to borrow from someone else’s case study.

What we checked before quoting any of this

The infographic reached us as a screenshot, and two of its figures did not survive a check against the report. One presents a link rate that the report measured only on mentions of Ahrefs itself as if it described brand mentions in general. The other, a traffic-to-signups conversion figure, does not appear in the report at all. Neither is in this article. We mention them because the check is the method: any figure that arrives through a graphic gets chased to its primary source before it goes into a client deck or a post.

What this does not tell a Quebec business

None of these datasets is Quebec, French-language, or local services. They measure the English-language commercial web, mostly American, weighted to software, agency and product verticals. The Quebec measurements we know of describe a different situation. The broadest published study (Observatoire Noos, étude #1) found local SMEs capturing 18 percent or less of model mentions in nine of twelve sector markets, and our own closed-book probe on Quebec roofing had the panel naming a verifiable local business in 7.1 percent of responses. Both are in the memory-layer post, with their caveats.

So carry the structure and leave the numbers. The five stages work in any market. The benchmark values for each stage have to be measured in yours.

What we do with it

Each stage has a check you can run yourself, most of them this month, and none of them is a promise of a result.

• Retrieved: confirm in your server logs that the named AI crawlers (OAI-SearchBot, PerplexityBot, Googlebot) actually fetch the pages that matter, and that robots.txt and your firewall let them through. The named crawlers, and which one governs citation on each platform, are laid out in the post on how AI chooses what to cite.
• Cited: run the questions your customers ask through ChatGPT, Perplexity and Google’s AI surfaces, and note which URL each one cites. Per platform, never averaged.
• Named: count mentions of your business against three named competitors, linked and unlinked as separate columns.
• Stay visible: repeat the same questions on a fixed schedule and log what changed. One reading is a snapshot; drift only shows up across several.
• Convert: in your analytics, isolate referrals from the AI assistants, break them out by landing page, and report conversion per visit.

That baseline, written down so the next quarter has something to compare against, is the opening move of an SEO audit sprint and the structure we report against in GEO work. It tells you which stage you fail before anyone spends money on the wrong one.

The headline says ranking does not decide who gets cited. The fine print says authority, other people’s pages and the platform you are measured on still decide most of it.