Competitors showing up in AI answers — and you're not?

When an AI engine keeps naming your competitor, that isn't a verdict on who does better work — it's a verdict on evidence. The engine assembled its answer from specific documents: roundup articles, review profiles, community threads, and pages on the competitor's own site. Nearly all of that material is public, and much of it is listed inside the answer as citations. So treat it as an investigation, not an insult: capture the answers, pull the citations, map the footprint behind their name, and sort their positions into entrenched and contestable. One evening covers it — the walkthrough is below.

Being recommended leaves a paper trail

An engine can know your competitor two ways: from its training data (a months-old snapshot of the public web) or from the live web search it runs mid-answer — the two pipelines explained in how ChatGPT recommends businesses. The live route is the one that matters here, because it shows its receipts: the sources it retrieved are either cited in the answer or one follow-up question away. Those citations are the closest thing this discipline has to discovery evidence.

The first thing the evidence usually shows: your competitor's website is the smallest part of their advantage.

One boundary before you start: this method is for the business that loses to a named rival. If the engines aren't naming anyone in your category — or don't seem to know you exist at all — diagnose the absence first with why your business doesn't show up in ChatGPT.

The investigation: five prompts, two engines, one evening

You need a spreadsheet, ChatGPT with web search, and Perplexity. To keep the steps concrete we'll follow Vista Dental Group — an invented Scottsdale practice from our example library, a fictional example and not anyone's real scan data — watching two rivals, Camelback Smile Studio and Desert Ridge Dental, take the Invisalign questions.

Step 1 — Capture the answers, every name in them

Write five buying-intent prompts the way a customer would type them, covering different question shapes: the category ask (“best invisalign dentist in scottsdale”), the qualified ask (“…for complex cases”), the cost ask (“how much does invisalign cost in scottsdale, and who is worth it”), the comparison ask (“camelback smile studio vs desert ridge dental”), and the trust ask (“is camelback smile studio good?”). Run each in both engines and record every business named in every answer.

Two engines is the minimum, not a nicety. Analyses of large prompt sets — Nightwatch's share-of-voice research among them — find the major engines agree on their answers less than half the time, so a rival who owns ChatGPT can be absent from Perplexity, and every disagreement between engines is a door standing open. In the Vista example, the ten captured answers (five prompts, two engines) name Camelback seven times, Desert Ridge four, and Vista once. That fraction of answer slots — their share of voice against yours — is the baseline number this whole exercise exists to move.

Step 2 — Pull the citations

Now expand the sources on each captured answer. Perplexity numbers its citations on every answer; ChatGPT shows source links when its search ran, and when it doesn't, a follow-up — “List the sources you used for that recommendation” — usually surfaces them. Paste each cited URL into your sheet next to its prompt. Expect a few dozen URLs with plenty of repeats; the repeats are the finding.

Some answers come back with no sources at all. That means the engine answered from model memory, not live search — there is nothing to expand, and that position shifts on model-release cycles rather than weeks. Note it, and check what the model itself believes about you with the five-prompt existence test in does ChatGPT know my business.

Step 3 — Map the footprint behind their name

Sort every citation into one of five buckets:

• Roundups and listicles— the “best X in Y” articles that answer engines lean on hardest.
• Review platforms — Google, Yelp, and whichever industry-specific sites the engines actually cited.
• Community threads — Reddit above all; why AI cites Reddit so heavily is its own story.
• Directories and profiles — the boring entity layer that corroborates a business exists.
• The competitor's own pages — usually the plain-spoken ones: cost, process, comparisons.

Tally citations per competitor per bucket, and note the date on every roundup. While you're at it, compare Bing footprints: search site:competitor.com and site:yourdomain.comon Bing. About 87% of ChatGPT's cited pages also rank in Bing's top results (Seer Interactive, 2026) — an overlap measured from the outside, not a pipeline OpenAI documents, since OpenAI names OAI-SearchBot as the crawler for ChatGPT Search and never says which index or indexes sit behind it. Treat a deep Bing footprint on their side and a thin one on yours as a strong lead rather than the verdict, and as a cheap thing to fix either way.

Step 4 — Re-run before you conclude anything

Run the same five prompts again a day or two later, and log who survived. This step is where most competitor panic dissolves:

A name that shows up across both engines and both runs marks real incumbency. A name that flickers was never entrenched — that answer is already in play, whatever it felt like the night you first saw it.

Reading the map: entrenched or contestable?

Score each question, not each competitor — the same rival can be immovable on one prompt and propped up by a single stale link on the next. A position reads as entrenched when the name survives both engines and re-runs, several fresh sources of different types back it, and their own site adds a page that answers the question plainly. It reads as contestable when one aging source does all the work, the engines disagree, the names change between runs, or the engine visibly stretched — citing a generic national page for a local question, or sources that never mention the question asked.

In the Vista example, Camelback's category question is entrenched: both engines, both runs, three current sources. But the cost question — the one that stung most — rests entirely on Camelback's own cost page. No roundup, no thread, one source. One source is one point of failure, and everything about Desert Ridge is more fragile still. The map turns a vague “they're everywhere” into a short list of specific, differently-defended positions.

Why their grip is looser than it looks

Even the entrenched positions sit on ground that gets re-tilled. The sources AI answers draw from are not a settled canon:

This cuts both ways, and honesty requires saying so: the positions you eventually win churn on the same schedule. That is why the investigation above is a cadence, not a ritual you perform once in a bad week — by hand it's worth repeating monthly, and the re-check, not the first look, is the method.

Counter-moves, matched to what carries them

The map dictates the move: attack the bucket doing the work, not the competitor.

• Roundups carry them. The exact articles are in your citation list. Pitch those publishers and the fresher lists ranking beneath them — writers refresh roundups, and stale entries are the ones that fall out.
• A community thread carries them. Join those communities honestly and by their rules — the compliant version of that work is in the Reddit citations playbook.
• Review platforms carry them. Concentrate your review cadence on the platforms the engines actually cited, not on every platform that exists.
• Their own page carries them. Publish your own plain answer to the same question — cost, process, comparison — shaped the way engines quote; the ChatGPT playbook covers the page shapes.
• Nothing carries anyone well. The open-net case: an answer-shaped page, correct schema, and a Bing submission is the fastest kind of win available in this discipline.

Sequence by winnability, not by pride. The positions to take first are contested, specific, long-tail questions — fragmented fields, stale sources — not your rival's fortress. Early wins compound, because every page that earns a citation becomes corroboration for the next one. And since none of the major engines sells placement in organic recommendations as of mid-2026, the footprint is the only door in — for them and for you. The full build sequence, ordered by time to impact, is in how to get your business mentioned by AI.

Our angle, so you can discount it

The company behind this guide, AEO Action, sells the automated version. The product runs the investigation above on a schedule instead of one evening: daily scans record who was named and which sources were cited for every buyer question, competitor pressure is tracked as its own sub-score rather than buried in an overall grade, and a Winnable Queries ranking scores every question you're missing on four measured components — how much its cited sources churn between scans, how tight the strongest competitor's grip is, how specific the question is, and how well it fits your business — from your own scan history, with each score stating what it's based on. A daily orchestrator then prepares one counter-move as a finished artifact — schema, an FAQ page, a landing page, or a Reddit pack — into an Approve Queue; nothing publishes without your approval, and each approved landing page is re-checked against its target question on following scans for up to 120 days, so you get a dated receipt: shipped on date A, first cited on date B. Running that loop by hand is what AEO agencies bill $2,000–6,000/month retainers to do (2026 market research); the scope of ours is written out on the AEO services page, and if you are comparing it against the enterprise category, the Profound comparison is the honest version. The four steps above are free and manual either way — this page works whether or not you ever run our scan.