It isn't often you get to look over someone's shoulder while they make a buying decision. That is exactly what the analysis published by Growth Memo in November 2026 offers, based on a thirteen-question Wynter survey of 200 people who bought software with an AI open beside them: their first messages, their follow-ups and what they did once they closed the chat.
This is GEO territory — optimising your website so AI search tools (ChatGPT, Perplexity, Gemini, Google's AI summaries) cite and recommend you — but with a twist that changes the priorities quite a bit.
What the survey says
Every figure below comes from that survey: 200 responses, all respondents from the SaaS/software sector.
- 158 of 200 (79%) arrived with at least one vendor in mind. Forty (20%) had none; two gave contradictory answers.
- Of those 158, 99 (62.7%) added a new vendor to their shortlist off the back of the conversation.
- Of the 40 who arrived with no candidates, 33 (82.5%) had one put on the shortlist by the AI. But only 40% of that group said they had chosen a vendor, against 69% of those who already had names.
- 78 of the 126 who chose (62%) say their pick already appeared in the first AI answer. Another 43 (34%) saw it appear after follow-ups. And 75 of those 78 kept asking questions anyway.
- Follow-up topics: pricing 57.5%, vendor comparisons 54.5%, integrations 46%, setup 40%. People with one or two follow-ups covered 2.4 topics on average; those with ten or more, 5.0.
- After the chat: 72.5% visited vendor websites, 48% went to Google, 31% asked peers, 29% searched Reddit and 23% checked G2 or other review sites.
- 67.5% used more than one AI tool. In this sample: Claude 71%, ChatGPT 62.5%, Gemini 38.5%, and Google's AI Overviews or AI Mode 15%.
- 68.5% started in a fresh chat with no extra context (only 52.5% picked that option on its own).
Why it matters if you sell software
The easy reading was that AI narrows the shortlist. The survey points the other way: it reopens it. Two thirds of those who already had candidates added one more. If you were the buyer's favourite, there is now a second door letting competitors in mid-process. If you weren't on the list, that door is yours.
Second: being cited first settles nothing. Seventy-five of the seventy-eight who ended up choosing the vendor from the first answer carried on interrogating the model afterwards. What decides it isn't the mention, it's surviving four follow-ups about pricing, comparisons, integrations and implementation. If your site doesn't say what you integrate with, what it takes to get running and how your pricing works, the model fills that gap with whatever it finds elsewhere: reviews, Reddit threads, third-party comparisons. Sometimes comparisons written by your competitor.
Third: the chat ends on your website. Almost three in four went to look at it afterwards. That page has to confirm what the model just said. If the AI claims you integrate with a specific tool and your site never mentions it, what breaks isn't your ranking — it's trust.
And fourth, measuring a single engine falls short when two thirds use several. On our self-service SEO platform, SeoTres, the «Does AI cite you?» tool does exactly that: it asks ChatGPT, Claude and Perplexity about your sector separately, so you can see whether the split in your niche looks anything like the one in your head. It usually doesn't.
What doesn't change
This is a self-reported survey of 200 people, with prompts reconstructed from memory and overlapping answer options. The analysis says so itself: it doesn't represent B2B buyers in general, it can't attribute an action to a specific tool and it establishes no causality. The fact that demo rates rise at six to ten follow-ups and fall again above ten doesn't prove that longer chats drive purchases; the case may simply have been harder.
Requesting a demo isn't buying, either, and the survey can't tell whether the vendor the AI introduced was the one finally picked. Watch the engine split too: Claude above ChatGPT says far more about the panel than about the Spanish market. If you allocate effort using those percentages, you'll get it wrong.
And no, classic SEO isn't dead: nearly half kept searching on Google after the chat.
Our take
The most useful thing in this study isn't a percentage, it's the shape: buying software has stopped being a funnel and now looks like a cross-examination. And in a cross-examination, the landing page full of adjectives doesn't win. The boring table does.
My bet, and you can argue with it: the obsession with «getting cited» will age badly. Mentions are cheap to measure, which is why everyone measures them. What pays is being recommended and surviving the follow-up. There's one example in the survey that sums it up better than any theory: a buyer asked the model to eliminate every vendor without an integration with a particular tool. The model needed a criterion for ruling people out, and applied it with whatever public information it had.
From which comes something counterintuitive: writing down what your product doesn't do leaves you better off. If your limits are clear and your competitor's aren't, the model rules out based on what it knows — and ends up ruling out the opaque one. Same with pricing: you don't need to publish rates, you need to explain how you charge.
Where I part company with the original piece is the closing jump to «build brand before the buyer opens a chat». True, but for a small company on a tight budget it's a blank cheque. The sensible order is the reverse: first fix what can already be cited — an integrations page with real names, technical requirements, an honest section on limits, accessible documentation — which costs hours rather than campaigns; then, with that standing up, spend off-site.
This looks a lot like what we run into when someone asks us to make AI speak well of their product: the problem is almost never the prompt or the engine, it's that the information needed to recommend it isn't written down anywhere public.
Any questions, tell me and we'll look at it. All the best, Vicente.