Buyers used to do their research in the open. They searched, clicked a few links, read a comparison page, maybe a review thread. You could see most of that journey in your analytics and your rank tracker, and you knew which pages you had to win.
That journey has moved somewhere you cannot see. A growing share of buyers now start inside an AI assistant, and more importantly, they stay there. The first answer is not the decision. It is the start of a conversation, and the conversation is where the sale happens.
Watch how people actually buy with ChatGPT, Claude, Gemini or Perplexity. They ask something broad: "what is the best tool for tracking my local rankings." They get three or four names. Then the real evaluation begins, in follow-ups: "is X good for a two-person agency," "X versus Y for a small budget," "what do people complain about with X," "does X integrate with my stack." Five, eight, sometimes twelve exchanges later, they have a shortlist of one. They never visited your comparison page. They never saw your retargeting ad. The whole middle of the funnel happened inside a chat window you have no access to.
This changes what "winning" means. On Google, you won a keyword. In an AI conversation, you win a position in an ongoing story the model is telling about your category. And that story is surprisingly stable. Assistants keep recommending the same few incumbents for months, because they learned the category from the same small set of sources: a couple of review sites, some directory listings, a Reddit thread, your own homepage if it is readable, and whatever else was well-structured when they ingested it.
Three patterns show up again and again when you track these conversations at scale.
First, being named is not the same as being recommended. Plenty of brands appear in an AI answer and still lose. The model mentions them in the first response, then talks them out of the deal in the follow-ups: "X is popular, but users say it gets expensive," or "X is powerful but probably overkill for you." If you only check whether your brand appears, you miss the part where the sale was lost.
Second, the follow-up questions are the real battleground. Brands optimize for the broad category query, but buyers rarely stop there. The questions that decide outcomes are specific: pricing, complaints, comparisons, fit for a niche use case. If the sources the model trusts say nothing about those, the model fills the gap with whatever it can find, which is often a competitor's comparison page or a three-year-old forum complaint you never answered.
Third, a small number of source pages do most of the talking. When you trace which pages the models cite and which pages their training data leaned on, it is usually fewer than ten. A stale directory listing with old pricing, a review roundup from last year, a help doc, your own homepage. Fixing those few pages moves AI answers faster than publishing anything new. I have seen an outdated listing describe a product's old free tier for months after the pricing changed, and every assistant repeated it as fact.
So what can you actually do? The measurement problem is real but solvable. You cannot get analytics from inside someone else's chat, but you can run the conversations yourself. Pick the ten to twenty questions your buyers ask, including the follow-ups, and ask them to the major assistants on a regular schedule. Write down who gets mentioned, who gets recommended after the follow-ups, and which sources get cited. That gives you a baseline and, over time, a trend line.
Then work the sources. For every answer you want to change, ask what page the model is relying on. If it cites a review site, get your listing there accurate and current. If it leans on a forum thread, answer the thread properly, as yourself, with your affiliation stated. If it quotes your own site, make sure the relevant facts are on a page the crawlers can actually read, in plain HTML, not buried behind JavaScript or blocked by a robots.txt rule you forgot about.
A word of caution on tactics. The obvious temptation is to flood the internet with AI-generated content praising your product. It does not work for long, and it burns the trust that makes AI visibility valuable in the first place. The models are getting better at discounting thin, repetitive promotion, and the communities that feed them, Reddit, forums, review sites, are ruthless with astroturfing. The durable play is boring: accurate, well-structured, genuinely useful information in the places the models already read, kept current as your product changes.
There is also a disclosure norm worth respecting. When you participate in the conversations and communities that shape AI answers, say who you are. A founder answering questions about their own product, clearly labeled, builds the kind of citation-worthy material models trust. A fake "happy user" doing the same thing is one screenshot away from a public correction that lives in the training data forever.
The uncomfortable truth is that no one fully controls what an AI assistant says about them. But "no control" is not "no visibility." You can watch the conversation, learn its patterns, and fix the inputs. The brands that treat AI answers as a channel to be measured and worked, rather than a mystery to be worried about, are already pulling ahead of the ones still only watching Google.
One more practical note: do this before you think you need it. AI answers shift slowly, which means by the time you notice a problem in your pipeline, the conversation has been saying the wrong thing for months. A lightweight monthly check of your category's key questions costs an afternoon. Repairing a wrong story that has hardened across every assistant costs quarters. The teams I see win at this are not doing anything exotic. They just started measuring earlier than everyone else.
I track this for my own product and for others, and the gap between the two groups gets wider every month.
Disclosure: I am the founder of Vouch (https://vouchgeo.com), a tool that tracks what AI assistants say about your brand across the full buying conversation.

