Answers
Why does ChatGPT recommend my competitor instead of me?
Short answer
Why does ChatGPT recommend my competitor instead of me?
An assistant recommending a product is not reading your website, it is reading a handful of third-party pages it retrieved, and your competitor is named on them. An assistant answering a recommendation question pulls a small set of sources, most of them not owned by any vendor in the answer, then writes from what those sources say. Across our own measurement of 901 citations spanning 82 distinct hosts, 91.5 percent of cited pages were off-site, which means the thing deciding the answer is almost never the page you control. A competitor who appears in the roundups, the comparison posts and the community threads the model retrieves will be recommended over a better product that appears in none of them, and no amount of editing your own homepage changes that.
The short version
901 citations across 82 hosts, our own measurement on our own money queries.
01
The answer is written from pages you do not own
When someone asks an assistant to recommend a tool, the model does not consult a ranking of vendors it holds internally. It runs a retrieval step, pulls a small set of pages that look like they answer the question, and composes from those. That set is the entire universe the answer is drawn from, so the practical question is not how good your product is, it is whether your name appears on the pages that got retrieved.
We measured which pages those are on our own money queries. Across 901 citations covering 82 distinct hosts, 91.5 percent pointed somewhere other than the vendor's own site. Roundups, comparison articles, forum threads and documentation belonging to other projects carried the answer. That number is our own measurement rather than an industry estimate, and it is the reason on-site work alone does not move a recommendation.
Our own measurement, 901 citations
901 citations across 82 distinct hosts on our own money queries. The surface that decides the answer is mostly not yours.
02
Being absent and being second look identical from your side
The frustrating part of this problem is that it gives you no signal. A page that ranks badly in a search engine still shows up somewhere on page three, and you can watch it climb. An answer either names you or it does not, and if it does not, nothing tells you whether you were narrowly beaten or never retrieved at all. Teams read that silence as a product problem and go and rebuild a landing page, which is the one intervention that cannot help.
The two cases need different work. If you are absent from the retrieved set, the job is getting onto the pages that get retrieved. If you are present but not chosen, the job is what those pages say about you. You cannot tell which you are looking at without measuring, which is why the first thing worth doing is a read rather than a fix.
What you observe, either way
Two different problems needing two different fixes, and one identical symptom. Telling them apart takes a read, not a guess.
03
Why one check tells you almost nothing
Before acting on what an assistant told you this morning, it is worth knowing how unstable a single reading is. In our step-zero run we asked 12 money queries five times each, 60 calls in total with every call succeeding, and 7 of the 12 queries returned a different outcome across identical repeats. Not different wording, a different answer to the question of whether a name appeared at all.
So a founder who checks once and sees a competitor has learned very little, and a founder who checks once and sees themselves has learned no more. Anything you intend to act on needs enough repeats to separate a real position from the coin flip, which is a measurement design question rather than a marketing one.
Step zero, 12 queries x 5 repeats
Our own run: 60 of 60 calls succeeded and 7 of the 12 queries returned a different outcome across identical repeats.
Asked next
The questions that follow this one
Can I make ChatGPT recommend my product by changing my website?
Rarely on its own. Our measurement puts 91.5 percent of cited pages off-site across 901 citations, so the pages carrying the answer usually belong to somebody else. On-site work matters for what a model says about you once you are retrieved, and it does not decide whether you are retrieved.
Is this different from ranking in Google?
The retrieval step rhymes with search, but the output does not. Search returns a list you can be tenth in. An answer names two or three products and omits everyone else, so the gap between third and fourth is the whole result rather than a position you can watch improve.
How fast does this change?
Faster than most teams expect. Ahrefs, measuring across 43,000 keywords, found AI Overviews persist about 2.15 days and 45.5 percent of citations change between consecutive observations, while the answer stays roughly 95 percent semantically identical. The sources rotate underneath a stable answer.
What is the first thing to do about it?
Measure which pages the assistant actually retrieves for your buying questions, and check whether you are absent from that set or present and unchosen. Those two diagnoses lead to different work, and guessing between them is how teams spend a quarter on the wrong one.
What to do next
Free gap report
When the model answers,
be the one it names
Tell us your category and the questions your buyers ask. We run them against live AI answers and walk you through what came back. If you are already winning, we will tell you that too.
12 queries · 5 repeats each · 5 working days
Free · Walked through live · Five working days
Queries we run
Cited instead of you
reddit.com38%
g2.com26%
news.ycombinator.com21%91.5% of citations point off-site