One Follow-Up Question Wipes 62% of AI Brand Picks — Measure Persistent Visibility

Key takeaway: A brand appearing once in an AI answer is a weaker signal than it looks. When Clovion analyzed 69,120 multi-turn conversations, adding a single buyer detail like "for a small team" made 62% of the brands in the first answer disappear from the second one (Search Engine Journal). Simply repeating the same question churned only about 10%. So this is not random noise — the ranking is rebuilt when the question gets specific. That is why one-shot "did we show up" checks mislead, and why you should measure across repeats, follow-up turns, and channel priority (as of 2026-07-08).

---

Three-line summary

• One appearance is not persistent visibility. Per Clovion, one buyer detail removed 62% of first-answer brands, while a plain repeat churned about 10%. Most of the movement is not random — it comes from the question getting specific.
• Studies with different methods point the same way. In the SparkToro/Gumshoe experiment, the odds of two responses returning the same brand list across 100 runs were under 1 in 100. A single-response snapshot cannot be trusted.
• So change the axis of measurement. Track not "did we rank" but "how often do we survive across repeats and follow-up questions" — and prioritize ChatGPT, which carries 92.4% of referral traffic, while watching for channel shifts.

The illusion of appearing once — one follow-up removes 62%

A brand showing up in an AI answer and that brand being recommended all the way through a buying journey are entirely different events. Clovion ran 69,120 multi-turn conversations across Claude, ChatGPT, and Gemini in 36 B2B software and fintech categories. It took the AI's first recommendation list as a baseline, then added the kind of one-line condition a buyer typically asks next and tracked how the list changed.

The result was clear. Adding one ordinary condition — "for a small team" — made 62% of the brands in the first answer vanish from the second one. Only about a third of the originally recommended list survived. Switching the condition to "for a large enterprise" produced a similar churn of about 72%. It was not the direction (small vs. enterprise) but the fact of getting specific that rebuilt the list.

One caveat comes with the number. Clovion is a vendor that sells AI visibility tracking, and this data is a self-study from its own methodology. The original release also carried an error — a dropped zero — so verified contradictions were 330, not 33, and tested brands were 2,040, not 204, after correction. That the sample actually grew after the fix adds confidence, but the single-vendor limitation remains. So this article does not rest on the 62% figure alone; below, it keeps only the points where a differently-designed independent study overlaps.

Not random noise — 10% on a repeat, 62% on specificity

The crucial distinction here is between "noise from repetition" and "rebuilding from specificity." That AI answers vary slightly each time is already known. An experiment by SparkToro's Rand Fishkin and Gumshoe's Patrick O'Donnell measured the size of it: 600 volunteers ran 12 prompts through ChatGPT, Claude, and Google AI a combined 2,961 times. The conclusion:

"There's a <1 in 100 chance that ChatGPT or Google's AI, if asked 100X, will give you the same list of brands in any two responses." — Rand Fishkin, SparkToro

Ask the same question 100 times and the odds of two responses being an identical brand list are under 1 in 100. That is the baseline noise from repetition. Yet in Clovion's data, churn from a plain repeat was only about 10%. In other words, "just ask again and 10% shifts" is the baseline, and "add a buyer condition and 62% changes" is a separate signal layered on top.

This difference is decisive in practice. The 10% of repeat noise averages out if you measure several times. But the 62% of specificity-driven rebuilding does not average away — it is a different answer to a different question in the first place. Real buyers do not ask "the best CRM" and stop. They always add conditions: "for a 50-person startup," "on a tight budget," "with Korean support." Conductor likewise found that intent type predicts the consistency of AI answers — three studies with different methods all pointing at the same thing: change the character of the question and the recommendation is rebuilt. Why conversational, long-tail queries should be the base unit of measurement is covered in content design for the conversational query era.

So which channel first — the reality of ChatGPT's 92.4%

Expanding measurement to repeats and follow-ups raises cost, which brings the "which channel first" question. Search Engine Land's analysis of 6.77M LLM sessions across 166 GA4 properties (November 2024–May 2026) found 92.4% of trackable AI referral traffic came from ChatGPT. The analysis frames the skew this way:

"Optimizing for 'AI visibility' without prioritizing ChatGPT means optimizing for an abstraction." — Search Engine Land

By measured traffic, that is fair. But two qualifiers attach to the number.

First, the channel mix itself is moving. In the same dataset, Claude grew 64x and overtook Perplexity in March 2026, while Perplexity fell 61% from its March 2025 peak; Gemini held a quiet second place with little volatility. Freezing your measurement to ChatGPT alone on today's 92.4% misses next quarter's reshuffle. Second, 28.8% of ChatGPT referrals landed on internal site-search result pages — a signal that the model trusts the domain but is unsure which page to send to, and a reason referral counts alone under-read ChatGPT's contribution. How to account for the indirect contribution of channels that rarely expose source links, like ChatGPT, is laid out in measuring AI traffic conversion and attribution.

In short, putting ChatGPT first is correct, but narrowing to "ChatGPT only" is risky. The reason to prioritize the dominant channel while still tracking the rising one is that citation rates and modes of contribution differ by platform — as documented in platform-specific AI visibility strategy.

So what should you measure — from one-shot exposure to persistent visibility

Where the three datasets overlap, there is one conclusion: "did we show up once" is not a metric but an illusion. To measure persistent visibility, shift the axes like this.

1. See repeats, not a single response. Even for the same query, two of 100 runs rarely produce the same list. Do not conclude "we're number one" from a single screenshot; judge on a repeatedly measured visibility rate.
2. See multi-turn, not just the first question. Real buyers add conditions. Measuring whether the brand survives follow-up context — "for a small team," "on a low budget," "with Korean support" — is how you find the 62%-churn zone. Look only at the first answer and you miss the half that disappears.
3. See priority, not every channel. Put ChatGPT, with 92.4% of traffic, first, but track fast-rising channels like Claude to get ahead of the next reshuffle.
4. Separate citation from mention. Being cited as a source and being named in the answer body are two different things. Miss the distinction and you mistake "cited but unnamed" for persistent visibility. The structure is covered in the ghost citation analysis.

On measurement, RanketAI's brand visibility analysis repeatedly measures whether a brand is cited and mentioned across ChatGPT, Perplexity, and Gemini answers over multiple queries and follow-up contexts, and competitor comparison shows the surfacing pattern against rival brands for the same query. Tools such as Semrush AI Visibility and Ahrefs Brand Radar also offer repeated, multi-prompt tracking — the point is to measure "persistent visibility across repeats and follow-up turns," not "one-shot exposure," with a tool that fits your market and language.

Frequently asked questions

Is it enough to check once whether our brand shows up in AI answers?

No. The SparkToro experiment concluded that even asked 100 times, the odds of two responses being the same list are under 1 in 100. A single check is just one point inside the noise, so you need a repeatedly measured appearance rate together with a survival rate when buyer conditions are added as follow-up questions.

Which should come first, repeat measurement or multi-turn measurement?

They serve different purposes. Repeat measurement filters out the baseline noise (about 10%) of "how stably do we appear in the first answer." Multi-turn measurement catches the rebuilding (up to 62%) of "do we survive when the buyer gets specific." With limited resources, confirm your baseline appearance rate through repeats first, then layer on a few core buying conditions as follow-up questions.

We have several channels — is it fine to watch only ChatGPT?

With 92.4% of measured referrals, ChatGPT is a fair first priority. But the mix is shifting fast — Claude grew 64x and overtook Perplexity in March 2026 — so freezing on ChatGPT alone misses the next reshuffle. Prioritize ChatGPT while also tracking rising channels.

Can we trust these numbers for the Korean market?

Clovion, SparkToro, and the referral data are all English-query based, and Clovion in particular is a vendor self-study. Rather than plugging the absolute values (62%, 92.4%, and so on) straight into Korean answers, use the structure — one appearance ≠ persistent visibility, specificity drives rebuilding, measure across repeats and multi-turn — as reference, and verify the actual values by measuring directly against your own brand and Korean queries.

Related reading

• AI search asks in 23 words, not 4 — content design for the conversational query era — why follow-ups and specificity are the base unit of measurement.
• Why AI visibility strategy must differ by platform — why citation rates and contribution modes differ by channel, and how to set priority.
• What is a ghost citation? 62% of AI citations end without a brand mention — the conceptual basis for measuring citation and mention separately.