Why AI recommends your competitor, and what it read first

An assistant doesn't have an opinion

Ask ChatGPT, Gemini, Perplexity or Claude which coffee roaster to buy from and you get an answer that sounds confident. It isn't guessing from taste. It is reading — reviews, local guides, forums, publications and your own website — and turning what it finds into a shortlist. Recommendation follows source presence and consistency, not brand preference.

That single mechanism explains most of the surprising results a brand sees. If a competitor is named more often, in more of the places an assistant reads, and named the same way every time, it keeps winning the answer — whether or not it actually makes the better product. There is no ranking algorithm to appeal to and no ad slot to buy. There is only the question of what the assistant found, and how much it trusted it.

That's a harder thing to accept than a search-engine ranking, because it means there is no single page you can perfect to fix it. The evidence an assistant weighs lives across dozens of pages you don't control, written by people who were reviewing a coffee shop, comparing software, or writing a local guide — not thinking about how a language model would read their sentence six months later.

Two kinds of knowing

It helps to separate what an assistant already believes about your brand from what it finds when it looks something up right now. Call the first model knowledge — the associations already built in from whatever it was trained on. Call the second live discovery — what a search-capable assistant reads today, at the moment someone asks. A brand can be strong on one and close to invisible on the other, and the two need different fixes.

A brand with strong model knowledge but weak live discovery is one that used to have a reputation but hasn't kept its public footprint current — old praise, no recent evidence. A brand with weak model knowledge but strong live discovery is newer, and everything it has going for it depends on what's freshly indexed right now, with nothing to fall back on if that goes quiet.

Assistants update their answers as they read new material, so a fix rarely repeats itself identically on the next question.

A coffee roaster with two very different scores

Say a small Portland coffee roaster — call it Northwind Coffee — is reasonably well known by name. Ask an assistant to compare “Northwind Coffee vs Blue Bottle” and it answers confidently, naming both. Ask the same assistant “best specialty coffee subscription” — a question that never mentions Northwind — and it recommends Blue Bottle and Stumptown, and stops there.

That gap is the whole story in miniature. Northwind wins the question that already contains its name, because the assistant only has to confirm what it half-remembers. It nearly disappears from the far more common kind of question — the one where a buyer is discovering, not confirming. Most of the buyers a small brand actually wants to win are asking the second kind of question, not the first.

Why the same three sources keep winning

Behind that leaderboard sit the sources the assistant actually reads: Google reviews, a local coffee guide, community threads, competitor pages. A rival that shows up in most of them, described the same way each time, becomes the safe answer for the assistant to give. A brand that is inconsistent — the wrong hours in one review, no page explaining where it roasts or ships — becomes a risk the assistant quietly avoids naming.

Consistency beats a single good page

This is the part that surprises people: one excellent homepage does not fix it. An assistant is weighing dozens of sources against each other. A single strong page competing against ten consistent mentions of a rival usually still loses. The fix has to be spread across the sources an assistant reads, not concentrated on your own site — which is a genuinely different kind of work from redesigning a landing page.

The same mechanism, whatever you sell

Swap the category and the pattern holds. A workflow-automation vendor can be strongly present when a buyer names a competitor directly, and absent from a generic “best automation platform for multi-location businesses” question, because the competitor has stronger use-case content and more third-party category coverage. A hotel can be well known locally and still absent from “best luxury hotel in Dubai for couples” because its destination and experience information is thin next to a competitor's, and its external authority is scattered. A property developer can be recognised by name and invisible in unbranded investment questions for the same reason. It's never really about the product. It's about which brand handed the assistant the clearest, most consistent evidence to work from.

“Can't I just ask ChatGPT myself and see?”

You can, and it's worth doing — the manual method is a genuine first step. What one person asking a handful of questions once can't give you is a reliable pattern: a probabilistic system answers differently from one minute to the next, so a single exchange proves very little on its own. The value shows up once the same questions get asked repeatedly, on a schedule, and the answers get compared to each other rather than read in isolation.

From noticing to fixing

Knowing this is happening is only the first step, and it's the step most visibility tools stop at. A useful version of this measures the same questions on a schedule, works out which sources are driving the gap, makes the changes that are safe to make on their own, and then checks whether anything actually moved by asking the same questions again. Measure, diagnose, fix, prove — in that order, on a loop, rather than a one-off report that hands you a to-do list and leaves the rest to you.

It's worth being precise about what “fixing it” doesn't mean, too. It doesn't mean writing more content about yourself, or repeating your name in more places on your own site — an assistant already has plenty of first-party material to draw on for most brands. It means closing the specific gaps in third-party evidence that are causing a specific competitor to win a specific question, which is a narrower and more answerable job than “improve our AI visibility” sounds like at first.

What actually moves the answer

1. 01Ask the questions your buyers actually type — not just your brand name.
2. 02Find out which sources the assistant is reading, and whether you appear in them at all.
3. 03Fix the facts at the source that publishes them, not only on your own site.
4. 04Ask the same questions again next month to see whether anything moved.

AI answers are probabilistic. VISIBLE reports measured movement and confidence — not guaranteed causality or ranking outcomes.

VISIBLE

None of this promises a fixed spot in an answer that changes by the week. It does mean the gap between winning your own name and winning the category question is measurable, and closing it starts with knowing which sources an assistant is reading about you today. That's what a free AI visibility check is for — see the full method if you want the mechanics first.