We followed 180 conversations to the third question
Does AI still recommend you after a buyer asks a follow-up? Most tracking asks one question and reads one answer. Buyers rarely stop there, so we measured the conversation instead.
Each conversation starts with a short question of two to four words, taken from what people search for in the brand's category. Two follow-ups come next: the buyer's situation (team size, country, the tool they use now), then the need that decides it. Each engine receives its own earlier answers with every turn, the way a chat does.
The sample: 90 conversations for 39 brands, none of them ours, each asked on Claude and Perplexity between October 17 and November 9, 2026. That makes 180 threads and 540 answers.
A third of the opening answers named no company
57 of the 180 first answers (31.7%) named no company at all. Asked a bare category, both engines usually explained it: what the service is, what to look for, questions to ask a vendor. Claude did this in 30 of 90 openers, Perplexity in 27.
Once the buyer added their situation, the empty answers fell to 30 (16.7%), and to 21 (11.7%) at the third question. A tracker that asks only the short question is often measuring an explainer, and reads it as "not named" for every brand in the category.
168 of 753 opening names were still there at turn three
The first answers named 753 companies in total. By the third answer, 168 of them were still named: 22.3% (19.5% to 25.4%). Of the 608 companies in the third answers, 72% hadn't appeared in the first.
In 116 threads, both the first and the third answer named companies. In 40 of them (34.5%), the two lists shared no name at all. The company named first in the opening answer was still named at the end in 43 of 123 threads (35.0%).
The lists also got shorter, from 4.2 companies per answer to 3.4. The follow-ups narrowed the field and swapped most of it.
The tracked brand arrived late more often than it dropped out
The brands themselves were rarely named on these questions, which is a finding of its own. Across all three turns, a brand was named in 20 of the 180 threads, covering 8 of the 39 brands.
Only 8 of those 20 threads named the brand in the first answer. In 12, it first appeared after a follow-up. In 5, a brand named in the first answer was gone by the third. These are small numbers, but they point one way: a check of the opening question would have missed most of the threads where the brand was recommended.
One long question didn't reproduce the conversation
Could you skip the thread and ask everything at once? For each topic we took its longest buyer question, asked on the same engine close to the same day. Where both answers named companies, 150 of the 599 names in the third turn (25.0%) also appeared in the long question's answer, and 68 of 157 pairs (43.3%) shared none.
The long question is a related question, not a word for word merge of the three turns, so some of that gap is the wording. Still, it can't stand in for the conversation.
An outside study found the same direction
Search Engine Journal reported in July 2026 on a study by Clovion, an AI visibility company, of 69,120 conversations on Claude, ChatGPT and Gemini across 36 B2B software and fintech categories. One ordinary follow-up, such as "for a small team", kept 28% of the recommended list. Asking the same question again kept 90%.
Different engines, categories and follow-ups, and a much larger sample. The direction is the same as ours: the buyer's details rewrite the list.
What we'd do
- Measure the third answer for the short questions buyers really type. The first answer is often an explainer or a long list nobody acts on.
- Write the follow-ups from your sales calls: who the buyer is, what they use now, the one need that decides it.
- Make that need easy to find on your own pages, in plain words: who the product is for, the constraint it handles, the country it serves.
- Treat a long, detailed question as its own measurement, not a shortcut for the conversation.
How Proofsource measures it
Proofsource asks a topic's short question and two follow-ups as one thread on engines whose API accepts the conversation's history, sending each turn the engine's own earlier answers, and records the companies every turn names. Each turn is stored as its own answer, so you can read what changed between them.
Where this could be wrong
The follow-ups were written by a model from each topic's buyer questions. Real buyers phrase them differently, and some would ask about price or support instead.
Two engines only. ChatGPT and Gemini aren't in this sample, and their numbers could differ.
Each thread ran once per engine. Some of the change is the ordinary drift between two answers to the same question, which we didn't measure separately here. Clovion's 90% on a repeated question suggests that drift is small next to the follow-up's effect.
Company names were extracted from the answer text, and a few are tools mentioned in passing, such as an integration. Names within a thread aren't independent, so the intervals are approximate. 39 brands chose to track these topics, and they may not look like your category.