What to do when AI recommends a competitor over you

Guide · AI Visibility · 6 min read · last verified 2026-07-24

The short answer: change the evidence, not the model

When an AI assistant keeps recommending a competitor instead of you, you cannot make it swap the name. What you can change is the evidence it reads before it answers. AI assistants assemble recommendations from the sources they retrieve at answer time — third-party listicles, review platforms, comparison pages, and a handful of authoritative domains — not from a fixed leaderboard you can petition. So the fix is a supply-chain problem: find the exact prompts where the competitor wins, pull the citations feeding those answers, diagnose why the competitor is present on those surfaces, then get corroborated onto the same surfaces and earn your own citations. This is slow, honest work. Nobody can guarantee AI will name you — the ranking is not fully observable — but you can measurably improve the odds by improving the underlying evidence.

Step 1 — Capture the exact prompts where you lose (measured)

Do not work from a vague sense that "AI likes them better." Write down the specific prompts. Buyers rarely type your brand name; they ask category and comparison questions like "best [category] tool for [segment]," "[competitor] alternatives," or "is [competitor] good for [use case]." Run each prompt across the assistants your buyers actually use — ChatGPT, Perplexity, Google's AI Overviews, Claude — and record verbatim: who gets named, in what order, and whether you appear at all.

Two facts make this step non-optional. First, answers are fragmented, not winner-take-all — the same prompt returns different names across assistants and even across reruns, so a single check is anecdote, not data. Second, according to industry analyses, AI Overviews now appear in roughly 45% of Google searches, meaning the answer box often replaces the ten blue links entirely. Log 15–30 real buyer prompts and mark each: you present / competitor present / neither. That frequency table is your baseline.

Step 2 — Pull the citations behind the answer (measured)

This is the step most teams skip, and it is the whole game. When an assistant names your competitor, look at what it cites. Perplexity and AI Overviews show source links directly; in ChatGPT with browsing, expand the citations. List every URL feeding the answers where the competitor wins.

You will almost always find the same pattern: the citations are not the competitor's own website. Industry analyses of AI citations consistently show brands get surfaced more through third-party sources than through their own domain, and that comparison articles take the largest single share of AI citations — roughly 33%. In practice the answer is being built from a "best tools" listicle, a G2 or Capterra category page, a Reddit thread, and maybe a Wikipedia entry. According to industry analyses, Wikipedia alone is about 7.8% of ChatGPT citations and Reddit around 1.8%,. Your competitor is not winning because their marketing site is better — they are winning because they are named and corroborated on the surfaces the model retrieves.

Step 3 — Diagnose why the competitor is there (derived)

Now group the citation URLs into the surfaces that actually decide the answer, and check whether you are present on each. This diagnosis is derived — you are inferring cause from the citation pattern, not reading the model's weights.

If a surface cites the competitor and never mentions you, that is a gap you can close. If it mentions you inaccurately, that is a correction you can pursue. Rank the surfaces by how often they appear across your losing prompts — fix the ones feeding the most answers first.

Step 4 — Act on the same surfaces, in priority order (derived)

You change the evidence by getting corroborated where the model already looks. Prioritize by expected impact, not by what is easiest. The Princeton GEO study (KDD 2024), which tested content changes across Perplexity, gives you a defensible ranking of what moves AI-answer visibility.

Translate that into moves: (1) Publish an honest comparison page — you versus the competitor — with real criteria, cited sources, statistics, and named quotes. Comparison content earns the largest share of AI citations, and it is one of the few surfaces you fully control. Never claim a win you cannot substantiate; models cross-check, and an inflated page gets ignored or contradicted by the corroborating sources around it. (2) Get corroborated on the third-party surfaces from Step 3 — earn listicle inclusion on real criteria, grow genuine reviews, complete review-platform profiles, and participate honestly in the communities that get cited. (3) Rewrite your own high-intent pages to the evidence pattern above: source every claim, add statistics with attribution, quote named experts, cut the keyword padding. This is where the Magrios thesis applies — measure the gap on each prompt, act on the specific surface feeding it, and re-measure. A source link should sit behind every claim you publish, because a source link is exactly what the model is looking for.

Step 5 — Re-measure on a locked benchmark (measured vs. hypothesis)

Corroboration takes weeks to be crawled, indexed, and pulled into answers, so you need a stable yardstick. Re-run the identical prompt set on a fixed cadence and compare against your Step 1 baseline. The benchmark must be locked — same prompts, same assistants, same scoring — or you are measuring your own methodology drift, not real movement. Because answers vary run to run, treat a single improved answer as a hypothesis and a sustained shift across repeated runs as a measured result. Expect partial, uneven progress: you may start appearing in Perplexity before ChatGPT, or in "alternatives" prompts before head-to-head ones. That is normal fragmentation, not failure.

Honest limits

AI ranking is not fully observable, and no one can force an assistant to name you or drop a competitor. Anyone promising guaranteed placement is selling something the mechanism cannot deliver. What is real and repeatable: the model answers from retrieved evidence, that evidence lives on surfaces you can influence, and improving your presence and corroboration on those surfaces improves your odds. Do the honest version — accurate comparisons, real reviews, sourced claims — because the dishonest version gets contradicted by the very sources the model reads alongside yours.