AI Names You, Then One Detail Erases You

You asked ChatGPT "best [your category]" last week, saw your name in the answer, and felt fine about it. That check was close to worthless. A real buyer never stops at the opening question, and the moment they add one ordinary detail, most of the brands named in that first answer are gone. This article proves that the follow-up, not the opening prompt, is where the sale is won or lost, and shows you what to write on your site so the model keeps you in the room.

The position, plainly: you are not optimizing to appear in the first AI answer. You are optimizing to survive the second question. Everything else is a vanity check that tells you what you want to hear.

Sixty-two percent of named brands vanish by the second question

The number comes from a study I made myself trace before I repeated it, and I will get to why that matters later. On July 7, 2026, Greg Jarboe reported in Search Engine Journal on new data from Clovion AI, an Oslo research group, drawn from a study they call "Surviving the AI Funnel." Clovion ran 69,120 multi-turn conversations across Claude, ChatGPT, and Gemini in 36 B2B software and fintech categories. Multi-turn means a real back-and-forth, the way a buyer actually talks, not a single question fired once.

Two findings sit at the center. When Clovion re-asked the exact same question a second time, roughly 90 percent of the originally recommended brand list stayed intact. The models were stable when nothing changed. Then they added one plain buyer detail, "for a small team," and only 28 percent of the first answer's brands survived. Sixty-two percent of the names dropped out. Not because the buyer rejected them. Because the buyer got specific, and the model quietly re-sorted the list toward brands that fit the new detail.

Translate that into money. If your name appears in the opening answer and you treat that as a win, you are counting a lead that evaporates the instant the buyer qualifies what they need. For a business closing even a handful of deals a month from AI referrals, being dropped at turn two is the difference between a full pipeline and a dry one, and you never see the drop happen because your one-time check said you were there.

You are checking the wrong question, and checking it once

Most AI visibility monitoring, including a fair number of paid tools, does one thing: it asks the model "best [category]" and reports whether your name shows up. Owners run that check, see themselves, and relax. I have watched this pattern across enough of the 300-plus businesses I have worked with to call it what it is. It is measuring the easy question because the easy question gives a comforting answer.

The problem is that "best physio in Manchester" is not how a buyer with a real problem talks. It is how a buyer starts. Within two or three messages they are typing "best physio in Manchester for sports injuries with evening appointments," and that is where the decision actually gets made. The opening prompt is throat-clearing. The qualifier is the buyer telling you exactly what they need, and that is precisely the moment the model reshuffles.

This is a different failure than the one I wrote about in how often AI names you across repeated identical asks. That piece is about consistency: ask the same question ten times and see how often your name survives run-to-run randomness. Useful, but a separate axis. What Clovion measured is not randomness at all. When the question was merely repeated, only about 10 percent of the list churned. When the detail changed, around 72 percent of the list swapped. Same brand list, same model, same moment. The only variable was specificity, and specificity is what knocked you out.

So the single-prompt check lies to you twice. It tests a question your buyers do not stop at, and it tests it in a way that hides the one thing that determines whether you get recommended: whether your fit is spelled out for the qualifier the buyer is about to add.

The list re-sorts by fit, not by chance

Clovion tested the direction of the churn on purpose. "For a small team" reshuffled the list. "For a large enterprise" reshuffled it by a near-identical margin, around 72 percent swapped either way, against roughly 10 percent when the question was just repeated. That symmetry is the tell. If the drops were random noise, the enterprise qualifier and the small-team qualifier would not produce matched churn. They do. The model is not rolling dice. It is re-sorting toward whichever brands state, in plain language a machine can read, that they serve that specific segment.

Hold this only as firmly as the evidence allows. Zahir Hasan, Clovion's COO, was careful about the causal claim. In his words, the link between how a model perceives your fit and whether it recommends you is:

"a strong, consistent coupling, not a proven causal law."

Zahir Hasan, COO of Clovion AI, quoted in Search Engine Journal, July 7, 2026.

That is a researcher refusing to oversell his own data, and it is exactly the honesty I want you to hold onto, because the fix I am about to give you follows a strong correlation, not a guaranteed lever. Hasan was just as plain about timing. Asked whether fixing this pays off in a set number of weeks, he said:

"We didn't do a before-and-after test. Treat it as worth testing, not guaranteed in X weeks."

Zahir Hasan, COO of Clovion AI, in the same Search Engine Journal report.

So I will not promise you a timeline. Anyone who does is selling. What I will tell you is that the mechanism is documented, the cost of ignoring it is real, and the fix is cheap enough that "worth testing" is more than enough reason to do it.

There is a second layer worth knowing, because it decides which fix comes first. Clovion found the models contradict each other on flat brand facts about 15 percent of the time, based on 330 verified contradictions across 2,040 brands. And the direction of the error depends on where your content lives. Claude and ChatGPT lean on documentation and product pages, and when a feature is poorly documented they hedge to "it doesn't have that." Gemini leans on marketing material and video, and tends to credit whatever you hype. In a separate July 7 piece, Search Engine Land's James Allen framed the two ways brands surface in AI answers, used versus cited, and the practical takeaway lines up: what the model can read about you decides what it says about you. Fix flat factual errors first. A model that thinks you lack a feature will drop you before fit ever enters the picture.

Write down who you are the right fit for, before a machine guesses

Picture your own business. You run a physiotherapy clinic in Manchester. Someone asks ChatGPT for the best physio in the city and it names you. Real win. Then they type "best physio in Manchester for sports injuries with evening appointments," and now it names three other clinics instead. Those three do sports rehab and open late, and so do you. The difference is that they wrote it down, in plain words, on a page the model reads, and you never did. That one missing sentence was the sale. Here is how to stop giving it away.

First, list the qualifiers your buyers actually add. Sit down and write the five to ten follow-up details a real customer tacks on. "For a small clinic." "On a tight budget." "Near me." "That integrates with QuickBooks." "For same-day appointments." You know these because you hear them on every intake call. This list is your test set, and it costs nothing but twenty minutes.

Second, run each qualifier through the models yourself. Open ChatGPT, Claude, and Gemini. Ask the opening question, confirm you appear, then add each qualifier one at a time and watch where you fall off. This is the specificity audit, and it is the honest version of the check you were doing. Do it by hand once so you feel the drop happen. If you want the disciplined how-to, my walkthrough on checking whether AI actually recommends your business lays out the prompts to use.

Third, write the missing fit statements in plain language. For every qualifier where you dropped off but genuinely fit, add a clear sentence to the relevant page. Not marketing fog. Machine-readable specifics: who you serve, what you specialize in, your hours, the tools you integrate with, your price range. "We treat sports injuries and offer evening appointments until 8pm on weekdays" is a sentence a model can lift and match. "Your recovery, reimagined" is not. The fix is not technical. It is writing the truth about your fit where a machine can find it.

Fourth, put those statements in the formats models trust. Product and service pages, FAQ blocks, and clearly labeled specs get read; a claim buried in a hero image or a PDF does not. Since Claude and ChatGPT lean on documentation, spell your specifics into text on the page. If you want to know which page types earn citations, I broke that down in my piece on the content formats AI actually cites. Match the format to the machine that reads it.

Fifth, if you delegate this, ask your agency one exact question. "Show me the follow-up qualifiers our buyers add, and show me where on the site each fit is stated in plain text." If they answer with rankings, impressions, or a "we're tracking AI visibility" dashboard that only shows the opening prompt, they are measuring the vanity check. Make them show you the second-turn results or find someone who will.

Track the follow-up, not the headline, or you will fool yourself

The vanity metric here is obvious once you name it: "we appear when someone asks for the best in our category." That number can be a clean 100 percent while you lose almost every qualified buyer at turn two. Stop reporting it as a win. Track the survival rate across your real qualifier list instead. Of the ten details your buyers actually add, how many keep you in the answer? That percentage is the number that maps to revenue, and it is the one to watch move.

Measure whole conversations, not single prompts. Re-run your qualifier audit monthly, because the models update and the list re-sorts when they do. Watch for the flat-fact contradictions too: if one model insists you lack a service you clearly offer, that is a documentation gap you can close this week, and it is dragging down every downstream answer.

Why does turn two carry so much weight? Because the AI answer is now the decision, not a signpost toward one. James Allen's Search Engine Land coverage carried the supporting numbers: Similarweb data shows AI-search traffic converting at 11.4 percent against 5.3 percent for organic, and Pew Research found in 2025 that users click a blue link only 8 percent of the time when an AI summary is present, versus 15 percent without. Read those together. Fewer people leave the answer to visit your site, and the ones the answer sends convert at more than double the rate. The recommendation itself is the moment of choice. Getting dropped from it at the exact instant the buyer qualifies is not a soft miss. It is the sale, closed for a competitor, invisibly.

The reason I made you wait for that 62 percent number

One more thing, and it is the reason I trust this study enough to build advice on it. The published Clovion PDF shipped with a dropped decimal. It read "204 brands" when the real figure was 2,040, and "33 contradictions" when it was actually 330. The Search Engine Journal column is, in part, a lesson in tracing a stat to its source before you bet a strategy on a headline. I have watched "SEO is dead" cycles come and go across 17 years in search, and every one of them ran on numbers nobody checked. So I check. Every stat in this piece traces to a named source, on a named date, because a strategy built on a misread figure is a strategy built on sand. That habit is the whole job.

Frequently Asked Questions

My business shows up when I ask AI for the best in my category. Isn't that enough?

No, and that check is the trap. Clovion's July 2026 study found that when a buyer added one ordinary detail like "for a small team," 62 percent of the brands named in the first answer disappeared. Your buyers never stop at the opening question. They add a qualifier, and that is where the model re-sorts the list. Test the follow-ups your customers actually type, not just the headline question.

Is fixing my site's fit statements guaranteed to keep me in AI answers?

It is not guaranteed, and I will not pretend otherwise. Clovion's COO Zahir Hasan called the link between perceived fit and recommendation "a strong, consistent coupling, not a proven causal law," and said they ran no before-and-after test, so treat it as worth testing rather than promised in any set number of weeks. The upside is that stating who you serve in plain language is cheap and low-risk. A strong correlation plus a near-zero cost makes it worth doing now.

What exactly should I write on my pages to survive the follow-up question?

Plain, specific fit statements a machine can read and match. Name who you serve, what you specialize in, your hours, the tools you integrate with, and your price range, in clear text on your service and FAQ pages. "We treat sports injuries and offer evening appointments until 8pm on weekdays" works because a model can lift it. Vague marketing lines do not. Fix any flat factual errors first, since a model that thinks you lack a feature drops you before fit even matters.

Go run the check honestly today. Ask for the best in your category, confirm you are there, then add the one detail your buyers always add and watch what happens. If you want a second set of eyes on where you fall off and what to write to stay in the answer, book a call with me and we will map your real qualifier list together. The uncomfortable question to sit with tonight is this: the last time you felt good about your AI visibility, which question did you actually ask, and would a real customer have stopped there?