When the answer names a competitor instead of you
The short answer: When an engine names a competitor instead of you, it almost never means it compared the two products and preferred theirs. It means your page was not in the candidate set it retrieved, or it was retrieved and nothing in it answered the question in a liftable sentence, or your name was there and the facts around it were too vague to use. Three different failures, three different fixes.
The reason this matters is that most teams read the answer as a judgement and respond by writing more marketing copy. That is the one response that reliably changes nothing.
The substitution happens before the model writes anything
An assistant with retrieval on does not hold a ranked list of vendors and pick one. It turns a question into one or more queries, fetches candidate documents, cuts them into passages, keeps a handful, and writes from what it kept. Google documents the expansion step for its own surfaces: its AI features guidance says AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics rather than one. Other engines publish much less about their retrieval, so for those we are reading the citation lists they show us and inferring the rest. The practical consequence, on the surfaces where we can check it, is that the competitor is named because their passage survived to the writing step and yours was not in the pile.
So the question to ask is not "why do they prefer my rival". It is "at which step did my page drop out". How an AI answer is assembled, step by step walks the loop. Here it is enough to know that the loop has four places to lose and the visible output tells you nothing about which one you lost at.
The three failures, and how they look different
You were never retrieved. The fan-out queries went to a category page, a comparison article, a directory and two review profiles. You are on none of them. Your own site may be flawless and it never entered the room. The tell is that your domain appears in zero citations across repeated runs of several related prompts, while your rival appears via third-party pages rather than their own.
You were retrieved and not usable. Your page was fetched and chopped, and every passage was a value proposition rather than an answer. The sentences we see quoted in answers tend to be ones that resolve the user's question on their own, which is a pattern from reading answers, not a documented selection rule. A paragraph that says you are the modern platform for growing teams resolves nothing. The tell here is that your domain does get cited, but on a different question than the one you care about, or it is cited while a rival's name occupies the actual sentence.
You were named and hedged. The engine mentions you in a list of also-rans with a thin, cautious description. This is usually an identity problem. Your pricing page, your directory listing and your about page disagree about what you sell, and the model takes the safest route, which is a short vague clause. Vague clauses lose to specific ones.
The two-run check that separates them
Run the same prompt twice on an engine that lets you control browsing: once with retrieval available, once with it switched off. The retrieval-off answer comes from what the model already absorbed during training, which is a slow, corroboration-heavy signal built from years of other people writing about the category. The retrieval-on answer comes from what was fetchable in the last few seconds.
Four outcomes, and each points somewhere specific:
- Rival named in both runs. You have a corroboration problem. The category's written
- Rival named only with retrieval on. Your page is losing a live fetch-and-parse contest.
- Rival named only with retrieval off. The record is stale and your current pages are
- You are named in both. Check the sentence next to your name before celebrating.
record does not include you, and on-site changes alone are unlikely to move it inside a quarter, because the off-site record has to change first.
This is the most fixable case and often the fastest.
working. Keep going, and get the third-party pages corrected.
That split is why audit prompt sets are run both ways. Our own set is about 45 prompts across five engines, run with retrieval on and off, and the pairs are more informative than either half alone. The sample audit report shows how the two columns sit next to each other.
Look at where the rival is cited, not that they are cited
Open the citations under an answer that named your competitor. Across the audit runs we have read, which use a set of about 45 prompts across five engines, the cited sources are more often third-party pages than the rival's own website: roundups, comparison pages, directories and review profiles. We have not counted that across a fixed, published sample, so take it as the pattern we keep seeing rather than a measured rate. It matches the pattern described in the third-party sources AI engines keep citing. That changes what you do next. Rewriting your homepage does not get you onto a list page you are absent from.
Two things move this. One is being on the pages the engine already fetches, which means claiming and completing profiles, getting into the comparison articles that rank for your category, and correcting the ones that describe you wrongly. The other is making the product facts easy to repeat: a named category, a stated price or pricing model, a clear statement of who it is not for. Review sites, G2 and what AI repeats about your product covers the profile side in detail, including the part most teams skip, which is reading what the profile currently says before asking for more reviews.
There is a trap in the opposite direction. Vendor-owned comparison pages come up rarely in the citation lists we read, and your own page about you and your rival is one of those. We cannot tell you why: none of the engine documentation we have read states a rule about interested parties, and the ranking reasons are not visible to anyone outside the engine. The page can still help, as the place that supplies correct facts for someone else to copy. Treat it as a source document for other writers, not as the thing that will be cited.
Numbers you can check, and what they do not tell you
Research on generative engines has shown that modifying a source document changes its chance of appearing in the generated answer. The KDD 2024 paper that introduced the term generative engine optimisation reported visibility gains of up to 40 percent from content-side changes alone, measured on GEO-bench, the query set the authors built for that paper. That is a research setting on a research benchmark, not a measurement of your site or of any live engine today, and the authors note the effect varies by domain.
The other number worth holding is about what a win is worth. Pew Research Center analysed 68,879 Google searches from 900 US adults in March 2025 and found that users clicked a result link on 8 percent of visits to pages with an AI summary, against 15 percent without, with the summary's own links clicked on 1 percent of visits. Being named in the answer is often the whole of the outcome. The click may never come, which is also why measuring this through traffic alone will mislead you.
Honest limits
We cannot tell you how many runs make a result real. Answers varied between identical prompts in our own runs, so a single run that names your rival is not evidence of anything, and we treat anything under five runs per prompt as noise. That five is our working convention, not a published threshold anyone has validated. The three failure modes above are a diagnostic frame we find useful, not a documented taxonomy: we can see that a page was not cited, but we cannot see the candidate set an engine assembled, so we are inferring where it dropped out. We also cannot connect a mention to revenue with any honesty, because the referrer data we see in client analytics from these surfaces is sparse, and we cannot measure what we are not sent. And we do not know how long a corroboration fix lasts, because the engines refresh their sources on schedules nobody publishes. If you want the cheap version of this diagnostic before spending anything, the self-check includes the two-run comparison.
The wider frame for all of it sits in what is answer engine optimisation.
FAQ
Does being named second in a list hurt me? Less than you would think. Order inside an AI answer moved between runs of the same prompt in our own testing, and the larger gap is between named and absent. Worry about the description attached to your name before you worry about its position.
Should I ask the engine why it chose my competitor? No. The model will produce a plausible rationale that is reconstructed after the fact, not a record of its retrieval. Read the citation list instead. That is real.
My rival is smaller than me and still gets named. How? Usually because the written record favours them. A small vendor with a complete directory profile, three roundup appearances and a specific one-line description beats a large one described in abstractions.
Will more blog posts fix this? Only for the second failure mode, where you are retrieved but unusable. If you are not being retrieved at all, publishing more of your own pages adds candidates that the fan-out queries still will not reach.
Sources
- Google Search Central, AI features and your website, including query fan-out: developers.google.com/search/docs/appearance/ai-features
- Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024: arxiv.org/abs/2311.09735
- Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results, 22 July 2025: www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/