Your main competitor shows up in AI answers three times more often than you do, and looking at them from the outside, you can't figure out what they're doing so differently. You're not losing on quality: you're losing because they have built precise signals on precise sources, and you don't yet know which ones. Analyzing their AI strategy systematically gives you a map of what to replicate, what to avoid, and where you can beat them.
Your competitor gets cited 3x more than you, but you don’t know what they’re doing differently. Here’s how to reverse engineer one of your rival’s AI strategy.
Has this happened to you? You open Perplexity, ask “best hazelnut producers in the Langhe for artisan pastry”, and the same three names always appear in the AI answers. You’re not there. And yet your product is good, you have a decent website, you’ve been doing SEO for years. What do the others have that you don’t?
A competitor AI audit exists precisely to answer this question. It’s not a classic SEO analysis: it’s an autopsy of the reasons why an AI model chooses to cite one brand and ignores another. Let me explain how it’s done, starting from a real case I worked on recently.
What a competitor audit looks like through an AI lens
A classic SEO audit tells you which keywords the competitor brings home, how many backlinks they have, how fast their site is. Important stuff, but the AI model only uses part of it.
An AI audit is different. It focuses on four dimensions that carry weight in generative answers: mention frequency (across how many queries the brand appears), mention quality (cited as a primary source, in passing, in a list), the sources that legitimize it (who talks about it on the web), linked entities (which concepts the model associates the brand with).
No academic paper explains this operational flow to you in a linear way: I infer it from the way RAG systems and models with web search retrieve and rank sources. It follows that, if you want to understand why the AI recommends your competitor, you have to take their presence apart the way the model sees it, not the way you see it on Google.
Why it sits upstream of all your GEO work
When I wrote about backlinks as a citation proxy and about implicit reference weight, I insisted on one point: AI models don’t only read your site, they read what the ecosystem says about you. The same goes for recognition of the author as an entity.
The competitor audit is the way to discover which ecosystem your competitor has built. It’s intelligence on the real playing field, not on the one you imagine. Without this step, any GEO strategy is blind: you’re optimizing blindly against someone whose moves you don’t know.
Below 25-30 queries you don’t have a sample: you have an anecdote.
The test you can run in 90 minutes
You need a spreadsheet, a list of 3-5 competitors (the ones the AI actually cites, not the ones you think it does), and access to ChatGPT, Perplexity, Gemini, Claude.
Here’s how to proceed:
- Build 30 realistic queries about your sector. For Langhe hazelnuts, examples: “best Tonda Gentile hazelnut producer for pastry chefs”, “difference between Piedmont PGI hazelnut and Turkish hazelnut”, “organic hazelnut supplier for the confectionery industry”, “Langhe hazelnut paste producers for artisan gelato shops”.
- Run each query on the four AI engines. For every answer note down: brands cited, position (1st, 2nd, in the final list), linked sources, sentiment (explicit recommendation / merely named / compared unfavorably).
- Collect the domain names of the cited sources. You’ll see patterns: an industry magazine, a consortium, a technical blog, a Wikipedia page.
- Check with the Google Knowledge Graph whether the top-3 competitors are recognized entities. Search for their name on Wikidata and on Google Business Profile.
At the end of the test you have a 30×5 matrix: 30 queries by 5 brands, with appearances, sources, context. Binary decision threshold: if a competitor appears in more than 12 queries out of 30, it’s dominant; below 5, it’s marginal; in between, it’s a challenger.
Identify the 2-3 recurring web sources that cite you little or not at all: that’s the first ground to work on.
The case I worked on in Nizza Monferrato
A Tonda Gentile hazelnut producer from the Langhe, in the Nizza Monferrato area, wrote to me in late 2025. He sold to about fifty artisan pastry shops and two coffee roasters in the North, with revenue under two million. He asks me: “Roberto, how come when I ask ChatGPT ‘Piedmont hazelnut suppliers’ three names always show up and I never do?”.
We ran the audit. The 30 queries, the four AI engines, the matrix. Result: the three competitors who systematically beat him didn’t have nicer websites than his. But they had three things in common. First: they were cited on the in-depth pages of the protection consortium and in two industry magazines with a clean web archive. Second: they had a Wikipedia entry or a Wikidata item associated with the company name. Third: the founder of two of the three was cited as a speaker at supply-chain events, with HTML pages that stayed online (not just program PDFs).
Limits of the test, stated upfront: a sample of 30 queries is not a statistical study, it’s an operational indication. Four AI engines at one precise point in time: tomorrow the pattern can change. But the signal was clear and consistent across all four engines, so it was credible.
The work that followed, over three months: a Wikidata item created for the company, two talks by the owner at supply-chain conferences with dedicated event pages, two editorial collaborations with industry magazines. When I re-ran the test in April 2026, the brand appeared in 9 queries out of 30. Not dominant, but present.
The mistakes I notice most often
Four patterns I find in almost every badly done audit.
Confusing “market competitor” with “AI competitor”. Your rival on the price list isn’t necessarily the one beating you in the AI answers. Often the AI competitor is a smaller producer but with a strong editorial presence.
Testing 5 queries and stopping. Below 25-30 queries you don’t have a sample: you have an anecdote. AI answers vary, you have to see the pattern.
Ignoring the cited sources. The sources the model links under the answer are the treasure map: they tell you where the model considers it authoritative to talk about the sector. Almost no one reads them.
Not comparing four AI engines. ChatGPT, Perplexity, Gemini and Claude draw from different indexes. If you rely on just one, you have a partial view. The rule: a brand cited on 3 engines out of 4 has distributed authority; one cited only by Perplexity lives on recent web search, not on consolidated knowledge.
What to do concretely this week
- Make the list of the 3-5 competitors that, in your opinion, the AI cites more than you. Write them on a sheet.
- Build 30 realistic queries for your sector (ask yourself: what would a customer looking for a supplier like me type?).
- Run the test on the four engines, record mentions and sources in a matrix.
- For the top-3 competitors, look for the Wikidata item, the presence in the sector’s editorial archive, the founder’s presence as an event speaker.
- Identify the 2-3 recurring web sources that cite you little or not at all: that’s the first ground to work on.
This is an entry-level check, it doesn’t replace a full analysis with professional tools (AI mention monitoring, brand tracking on LLMs, knowledge graph audit). But in one afternoon it gives you a snapshot of where you stand compared to whoever wins the AI answers in your sector.
Where we go from here
A well-done competitor AI audit doesn’t hand you visibility: it tells you exactly where to earn it. It’s the foundation on which you build your own strategy for showing up in your customers’ AI answers, not the abstract version of someone else’s.
In the next articles in this series on measurement we’ll see how to track your AI citation share over time, how to structure a brand mention dashboard on generative engines, and how to turn a one-shot audit into a continuous monitoring system. If you’re interested in the entity side, go back to the piece dedicated to event entity speaking authority: the Nizza Monferrato case moved on exactly that lever.
