Everyone in SEO is now worried about the same thing: what happens when an AI answers the question instead of sending someone to your website.
The hopeful version goes like this. The model reads your page, recommends you by name, links to you as its source, and you get the visit anyway. Being cited is the new being ranked.
We tested it. We asked six buyer questions about three well-known products across five AI platforms, and counted two things separately: how often each brand was named, and how often the answer linked back to that brand's own site.
The brands were named 33 times.
They were linked to zero times.
Results at a glance
A mention and a citation are different things, and only one of them can send you a visitor.
All three brands are winning the first and getting nothing from the second.
What we asked, and of whom
We used our own AI Visibility Checker, which writes six questions a buyer would ask before they know your brand exists. Questions containing the brand name are refused outright, because handing a model your name guarantees it repeats it back and measures nothing.
For Mailchimp, the tool read the site, worked out the category, and asked things like "What are the best options for Email SMS Marketing Platform?" and "What alternatives should I compare?" Ordinary buyer questions, with no brand in them.
Those six questions went to ChatGPT, ChatGPT Search, Claude, Gemini, Gemini Search and Perplexity. Then we counted mentions and citations separately.
Named 33 times, linked zero
| Brand | Answers that named it | Platforms | Links back |
|---|---|---|---|
| Zapier | 16 of 24 | 4 of 6 | 0 |
| Mailchimp | 11 of 30 | 5 of 6 | 0 |
| Trello | 6 of 24 | 2 of 6 | 0 |
Three products of very different sizes, in three different categories, on the same day. The mention counts differ enormously. The citation count does not differ at all.
The line at the bottom of every report
Each of the three reports ended the same way, in the panel headed Your pages the AI cited:
"None of your pages were cited in these answers. The AI is answering this category from the sites listed here instead."
That last clause is the part worth sitting with. The answers were not invented. The models did read sources — Gemini Search cited six for Mailchimp and four for Trello.
Just none of them belonged to the brand being recommended.
So the model recommends Mailchimp on the strength of what other people wrote about Mailchimp. The company's own website, which it spent twenty years and an enormous amount of money building, was not part of the answer.
Why a mention is not a visit
It is worth being precise about what each of these is worth, because they are often lumped together as "AI visibility".
A mention is reputation. Someone asked which email tool to use and the model said Mailchimp. That is real and it matters. It is also unmeasurable in your analytics, unattributable in your reporting, and impossible to prove to a board.
A citation is a visit. A link the reader can click, a referrer in your logs, a session you can count.
Search rankings gave you both at once, which is why nobody had to separate them for twenty years. AI answers give you the first and, on this evidence, frequently not the second.
If your plan for AI search is "get mentioned and the traffic follows", the traffic may not follow.
What actually earned the mention
Here is the useful part, and it points somewhere unglamorous.
If the models are answering from other people's pages, then being written about elsewhere is what puts you in the answer. Not your homepage copy. Not a file at the root of your domain — we tested that separately and Google's own documentation says it does nothing for search.
Reviews, comparisons, forum threads, roundups, documentation on sites you do not own. The same surfaces that earn links are the surfaces that feed answers, which means the work has not changed as much as the panic suggests.
What has changed is the payoff. That work used to bring a link and a visitor. Now it may bring a mention and nothing else — so it needs to be worth doing for the reputation alone, and you should stop modelling it as a traffic channel until you can see traffic.
Copy this in an afternoon
Four steps, none of which need a budget.
1. Measure the two things separately. Mentions and citations are different numbers with different value. Any tool or report that blends them into one "visibility score" is hiding the finding above.
2. Ask the questions a stranger would ask. Never put your brand in the prompt. "Is Mailchimp good?" gets you named every time and tells you nothing. "What are the best options for email marketing?" is the real test.
3. Find out which sites the models are reading. Every report lists them. Those pages are shaping your category whether you like it or not, and they are where your effort belongs.
4. Stop forecasting AI traffic you cannot see. Report mentions as reputation and citations as traffic, and never let the first be presented as the second.
What we could not measure
Three brands on one day is an illustration, not a law. Six questions each, one run apiece. A different category, a different day or a different set of questions would give different mention counts. What would surprise us is a different citation count, but we have not shown that.
Claude was not reached in any of the three runs, and several other legs did not return on individual runs. So the platform coverage is uneven, and the totals count only the answers we actually received.
This is not what a customer sees in every case. Some legs search the live web and are the only source of citations; others answer from training memory with no search at all. The tool separates them and never averages them, and neither should you.
We cannot see any brand's real referral traffic. Zero citations in 33 answers is not proof that these companies get no AI traffic at all. It is evidence that the answers we sampled sent none.
Our AI Visibility Checker asks six buyer questions across the major assistants and reports both numbers apart: which answers named you, which linked to you, and which sites the models read instead. It refuses any question containing your brand, because that would guarantee a mention and measure nothing.
- Mentions and citations counted separately, never blended into a score
- The sites the AI actually cited, so you know who is shaping your category
- 3 checks a day, no signup, no card
Frequently asked questions
What exactly did you run?
On 23 August 2026 we ran Zapier, Mailchimp and Trello through our AI Visibility Checker. Each check writes six buyer questions from the brand's own site, refuses any question containing the brand name, and puts them to ChatGPT, ChatGPT Search, Claude, Gemini, Gemini Search and Perplexity. Zapier was named in 16 of 24 answers across 4 platforms, Mailchimp in 11 of 30 across 5, and Trello in 6 of 24 across 2. None of the three received a single link.
What is the difference between a mention and a citation?
A mention means an answer named the brand. A citation means the answer linked to that brand's own site as a source. Only the second one can appear in your analytics or send you a customer, which is why they are counted separately and never combined into a single score.
Did the models cite anything at all?
Yes. The searching legs cited sources — six for Mailchimp, four for Trello — so the answers were grounded in real pages. None of those pages belonged to the brand being recommended. Every report ended with the same line: none of your pages were cited in these answers.
Does this mean AI sends no traffic?
No, and we are careful not to claim that. It means the 33 answers we sampled sent none. We cannot see anyone's real referral logs. What we can say is that being named frequently, as all three brands were, did not by itself produce a single clickable link in this sample.
Why was Claude not reached?
It did not return in any of the three runs, and several other legs failed on individual runs too. The totals count only answers we actually received, which is why the denominators differ between brands — 24 for Zapier and Trello, 30 for Mailchimp. Uneven coverage is a real limitation of a single free run.
Would an llms.txt file fix this?
There is no evidence it would. We surveyed adoption separately and found that Google's own guidance tells site owners they do not need machine-readable AI files, that Search does not use them, and that creating them neither helps nor harms. Zapier publishes one and still received zero citations here.
So what does put you in the answer?
On this evidence, other people's pages. The models cited real sources, just not the brands' own. That points at reviews, comparisons, roundups and forum threads — the same surfaces that have always earned links. The work is familiar; what has changed is that the payoff may now be reputation without the visit.
Can I run this on my own brand?
Yes, free, three checks a day and no account. Enter your brand and domain and the questions are written for you, or write your own — with the one rule that they cannot contain your brand name, because a model handed your name will simply repeat it back.
Count the two numbers apart
Most of this series has been about what companies built. RTINGS built a measurement lab. wikiHow drew a picture for every step. Goodreads found sixteen nouns where everyone else found one. IKEA moved its hreflang out of the page.
This one is about something nobody built, and everybody assumed: that being recommended by an AI would work like being ranked by a search engine.
Thirty-three times, three well-known brands were exactly what the model recommended. Not once did the answer point at their website.
Being named is worth something. It is just not worth what a ranking was worth, and the gap between those two things is the number your reporting should be showing.
So run the check on your own brand and look at the two figures side by side. One of them is reputation. The other one is traffic.
Measured on 23 August 2026 using our own AI Visibility Checker: three brands, six buyer questions each, put to six assistants, with mentions and citations counted separately. Per-brand and per-platform figures are in the FAQ above.
Three brands on one day is an illustration rather than a law, platform coverage was uneven, and no claim is made about any company's actual referral traffic, which we cannot see.