What to Do When a Competitor Owns the AI Answer for Your Category

You type your category into ChatGPT the way a buyer would, and the same competitor's name comes back every time. Perplexity does it too, with three citations behind it. You know your product is better on the axes that matter, and none of that reaches the answer, because the answer was assembled before your product was ever in the running. The reflex is to assume the model has an opinion about you. It does not have one yet. What it has is a retrieval pipeline that keeps handing it your competitor's name and never handing it yours, and that pipeline is a set of mechanical stages you can inspect and move.

This is a bottom-funnel problem wearing a top-funnel disguise. The person losing sleep over it is close to a purchase decision, watching a specific rival collect the recommendation that should be a coin flip. Fixing it is slow and it favors the incumbent, so anyone promising a two-week reversal is selling something. What follows is a staged plan built on the AI Visibility Funnel, the four steps a query passes through on its way to naming somebody: Retrievability, Candidacy, Citation, Attribution. You work the stages in order, because a gain at the citation stage is wasted if you never entered the candidate set in the first place.

First, split the loss into retrieval and naming

Before you change anything, find out which stage you are losing at, because the two failures call for opposite work. Either your competitor sits in more of the candidate sets the model retrieves, or you both get retrieved and the model names them more often once you are both on the table. These look identical from the outside and they are not the same problem.

Run the diagnosis by hand first. Take fifteen prompts a buyer would actually ask in your category and submit each one several times to ChatGPT and Perplexity, since those two expose their sources most readily. For every answer, record two things: whether your competitor is cited or named, and whether your own pages appear anywhere in the source list even when your brand is absent from the prose. A competitor cited in the sources on twelve of fifteen prompts while you appear in zero source lists is a retrieval loss. A competitor and you both showing up in the sources, with only their name making it into the written answer, is a naming loss. The Princeton GEO research is about the second kind, since it measures what raises visibility once a source is already in the context window, and it is the wrong tool if your pages never reach the context window at all.

Most brands who feel invisible are losing at retrieval, not naming. That is the harder loss and also the more fixable one, because retrieval runs on inputs you can enumerate.

Map the citation surface your competitor is riding

If the loss is retrieval, the next question is which pages are carrying your competitor into all those candidate sets. Attribution is not the same as retrieval, and the URLs a model cites are your best available window into what it retrieved, so treat the citation list as evidence about the layer underneath.

Go back to your fifteen prompts and pull every URL cited across every run, then set aside the ones that mention your competitor. You are building their Citation Surface, the full set of third-party pages where an engine can find and cite them for your category. It will cluster fast. A handful of review-site category pages, two or three comparison articles, a Reddit thread or two, maybe an analyst roundup, and those same domains will keep reappearing across different prompts. Rank them by how many of your prompts each one gets cited for. The page cited for eleven of fifteen prompts is load-bearing for your rival's whole presence in the category. The blog post that surfaced once is not.

One number matters more than any single page here, and it reframes the whole effort. Category studies frequently find that third-party pages contribute materially to brand mentions, although the ratio is not universal. Measure it directly for your market: the pages deciding this are mostly pages your competitor does not own either. They earned their way onto them, and so can you.

Find the retrieval gap where neither of you is winning

The ranked surface tells you where your competitor lives. The more valuable map is the negative space around it, the prompts where the incumbent is weak or absent, because contested ground is cheaper to take than occupied ground.

The Retrieval Gap is the distance between all the pages that could answer a query in your category and the smaller set the engine actually retrieves. Two kinds of gap are worth hunting. The first is prompts where your competitor gets named but on thin evidence, cited to a single generic page, which means the model is reaching because it has nothing better and a specific, well-sourced page could displace the generic one. The second is prompts where neither of you appears, where the answer names two other players or stays vague, which is open territory nobody has claimed. Run a wider prompt sweep, thirty or forty prompts across your intent clusters, and tag each one by who wins and how strong the evidence behind them looks. The weak-win and no-win prompts are your entry points. Attacking the prompts where your rival is cited five times over on authoritative pages is the slow, expensive front, and you fight it last.

This is Prompt-Space Coverage doing its real job. A flat keyword list would tell you the category is competitive and stop there. Mapping the actual space of prompts and your visibility across each region shows you the soft spots, and the soft spots are where an entrant with a time disadvantage should spend first.

Close the candidacy gaps that keep you out of the running

Now you act, and the first target is Candidacy: getting your content into the candidate sets at all. Three things gate this, and they are unglamorous.

Coverage of the customer prompt-space comes first. For every intent cluster where you are absent, you need content that genuinely answers that prompt, written to match how buyers phrase it rather than how your marketing team phrases it. Self-contained passages are easier for a retrieval system to score and reuse, so structure the answer so a single retrievable chunk resolves the question without needing the rest of the page. A prompt like "what's a good alternative to [incumbent] that doesn't get expensive at scale" needs a page that answers exactly that, not a homepage that gestures at it.

Presence on the third-party pages comes second, and it is the bigger lever given where mentions actually come from. Take the citation surface you mapped and find every page where your competitor appears and you do not. Some you can influence directly: claim and update your profile on the review platforms, submit to the directories and listicles that accept vendor entries, correct the comparison pages that are simply wrong or outdated about you, answer the real question in the forum thread under your own name. Others you can only earn, by being worth the mention, and those are slower and more durable. Fix the absences you can fix this week, then start the long work on the ones you have to earn.

The third gate is the plain one. If the AI crawlers cannot fetch and index your pages, none of the above enters retrieval, and this is where a surprising number of brands quietly lose. Check your robots.txt against the real user-agents: GPTBot, OAI-SearchBot, and ChatGPT-User for OpenAI, ClaudeBot and Claude-User for Anthropic, PerplexityBot and Perplexity-User for Perplexity, plus Google-Extended and CCBot. A blanket disallow or firewall rule can exclude search-specific crawlers from the index an assistant retrieves against. Treat training crawlers separately, because blocking training does not necessarily block live search. Confirm the pages are being fetched and are present in the index before you attribute a loss to anything subtler.

Win the citation stage by being the specific source

Once you are reliably in the candidate sets, the fight moves to the citation stage, where the model chooses whose name to write. Here the Princeton finding earns its keep. When researchers tested what raises a source's visibility in generated answers, the methods that worked were adding citations, adding statistics, adding direct quotations, and improving fluency, with lifts reaching around 40 percent on their visibility metric. Keyword stuffing did nothing or slightly hurt. Lower-ranked sources gained the most, some over 100 percent, so this stage rewards the challenger more than the incumbent.

The practical move is to be the specific, quotable source for claims where your competitor is generic. If the load-bearing page recommending your rival says they are "a popular choice for growing teams," and your page says teams above 50 seats cut onboarding time by a measured amount with a named mechanism and a source behind it, the model has a reason to reach for your passage when a prompt gets specific. Information gain is the lever. A page that adds a fact the rest of the surface does not carry is the page a reranker keeps surfacing, because it resolves questions the generic pages leave open. Effects here are domain-dependent, so weight your effort by category: citations help most on factual queries, an authoritative tone on debate and history topics, statistics on law, government, and opinion. Match the evidence type to what your buyers actually ask.

Measure it like an experiment, not a screenshot

The failure mode of all this work is judging it from a single screenshot on a good day. Retrieval is not deterministic, so one run tells you what happened that time and not what tends to happen.

Fix a prompt set and a schedule before you start changing pages, so you have a baseline. Run each prompt multiple times per engine, because a brand that appears in one of five runs and a brand that appears in five of five are in very different positions, and a single query cannot tell them apart. Track Share of Retrieval, the fraction of candidate sets your content lands in, as your leading indicator, since it moves before citations do and tells you the candidacy work is landing even while the named-answer rate still lags. Watch the naming rate as the lagging confirmation. Expect the retrieval number to move first by weeks, because the model has to re-crawl, re-index, and start selecting your pages before any of it reaches the written answer. Doing this capture and sampling by hand across four engines twice a month is exactly the task that gets dropped the first busy week, and tools like shareof.ai↗ exist to keep the measurement running so you are steering on a trend rather than a lucky query.

None of this dislodges a well-entrenched incumbent quickly, and you should plan for months on the prompts they own outright. What moves sooner is the contested ground: the weak-evidence wins and the open prompts where you were simply never a candidate. Take those first, measure the retrieval share climbing, and let the harder territory come as the crawlers catch up to the work you have already shipped.

Primary sources and evidence

• GEO: Generative Engine Optimization (Aggarwal et al.)↗ — the original paper and benchmark.
• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks↗ — the foundational RAG paper.
• Web search in the OpenAI API↗ — current search-tool behavior and citations.
• OpenAI web crawlers↗ — official user-agent controls for search and training.
• OpenAI web crawlers↗ — GPTBot, OAI-SearchBot, and ChatGPT-User controls.
• Perplexity crawler documentation↗ — PerplexityBot and Perplexity-User guidance.
• Google common crawlers↗ — Googlebot and Google-Extended distinctions.

Continue through the system

• Map Your Citation Surface: Every Page Where AI Can Find You
• Reverse-Engineering Co-Citation: Find the Pages That Decide Your Category
• A/B Testing Content for AI Citation: A Methodology That Survives the Noise

Run the practical layer with the free AI visibility scan, explore the AI search library, or see the AI visibility benchmarks.