The competitor the AI cites more often than you hasn't found a secret formula: it has built precise signals on precise sources, and those signals are readable. Every citation it receives is a map of what the AI considers trustworthy in your industry. Reconstructing that map starting from its mentions gives you an operational list of what to do — instead of proceeding by trial and error.
Take the competitor the AI cites most in your industry. Open ChatGPT, run five different queries about your market, note which brand keeps coming back as a source. Now reverse engineer its mentions: which publications it appears on, with what sentiment, with which recurring claim. That pattern is your map, not an opinion.
Let me say it right away: this is not an article about “spy on your competitor”. It’s an article about how to reconstruct the mental model that ChatGPT, Claude, Gemini and Perplexity have of your industry starting from the brands that are already inside it. If the competitor is there, it’s there because it has accumulated specific signals — publications, formats, data, schema. Reconstructing those signals is the closest thing to an objective audit of AI visibility that you can do without enterprise tools.
What it means to reverse engineer a competitor cited by the AI
In the previous articles in this series I showed you how to measure your presence in AI answers, how to track citations, how to read the sentiment that emerges around your brand. Reverse engineering the competitor starts one step earlier: instead of asking yourself why you’re not there, you ask yourself why they are.
In the world of research on generative models, the documented mechanism is that the model retrieves fragments from a narrow set of sources that have passed a series of authority filters — recognized-entity signals, density of co-occurrences on vertical publications, readable semantic structure. From this follows an operational consequence that matters for your business: if a competitor passes those filters and you don’t, the difference is almost never “the product”, it’s the fabric of signals around the brand.
And that fabric is inspectable. Not perfectly, not with one click, but enough to derive a concrete action plan from it.
Why this work comes upstream of any AI-oriented content strategy
If you don’t know what makes the competitor the AI chooses work, you end up producing content on instinct. And content on instinct, in AI visibility, is the fastest way to stay invisible.
The pattern of the cited competitor is the synthesis of years of choices: which publications talked about it, in what format, with what data. When ChatGPT generates an answer about your industry, it’s drawing from that synthesis. Understanding the pattern means understanding which form of content the model considers credible for your category.
It’s work that builds on everything I’ve already told you about author entity recognition and implicit reference weight: signals that AI engines use to decide who deserves to be a source.
Mistake two: copying the format without copying the level of depth.
The hands-on test I ran on the high-end sports footwear makers of Civitanova Marche
Let me tell you about an indicative test, not a formal study. Civitanova Marche and the MC area are one of the historic hubs of high-end sports footwear in Italy: five or six manufacturers that work for the leading international brands and have their own premium lines.
I opened Perplexity and ChatGPT and ran six queries of the type: “best Italian makers of premium made-in-Italy sneakers”, “high-end sports footwear makers Marche”, “who makes luxury handcrafted sneakers in Italy”, plus three English variants on the same theme. I recorded for each answer.
- which brands from the district were named
- which sources the model drew from (publications cited under the answers)
- which claim recurred about the most-mentioned brand
Out of six queries, four named the same district manufacturer first. The sources cited by the models concentrated on three vertical publications (one fashion-business, one footwear-industry, one international premium lifestyle) and on two short encyclopedic entries that treated it as a recognized entity. The recurring claim was always the same: Marche manufacturing tradition + collaborations with well-known designers + top price tier.
Limits of the test: small sample, two engines, an afternoon of work. But the pattern was so clean that you didn’t need a larger dataset to read its meaning.
Choose eight queries that represent the need of your typical customer.
How to do the reverse engineering step by step
Now let me explain the process I followed, so that you can replicate it in your industry in two or three hours of clean work.
First step: identify the most-cited competitor. Open ChatGPT and Perplexity, run five to eight queries representative of your market (not of your brand: of the customer’s need). Count the mentions. The brand that comes up most often is your reference point.
Second step: extract the list of sources. On Perplexity the sources are visible below the answer. On ChatGPT with search active they’re linked inline. Save them all. On average you’ll find 8-15 of them for the main competitor.
Third step: classify the sources by type. How many vertical industry publications? How many authoritative generalist ones? How many of the competitor’s product pages? How many third-party pages (retailers, reviews, rankings)? How many encyclopedic or knowledge-graph entries?
Fourth step: read three of the source pages and look for the pattern. Average length, presence of numerical data, schema markup (checkable with Google’s Rich Results Test by pasting in the URL), type of heading, presence of direct quotes from the founder or from company figures.
Fifth step: check whether the competitor has a Wikidata entry. Go to Wikidata and search for the brand: if it exists as an entity with populated properties, that’s one of the reasons the model recognizes it with confidence. A related topic I’ve already covered in the article on the Google Knowledge Graph entry.
This is a first entry-level step: a real analysis, across dozens of queries and with longitudinal tracking, requires professional tools. But it already gives you 70% of the signal you need to understand where to move.
The mistakes I see most often when companies try to do this work
Mistake one: confusing “market competitor” with “AI competitor”. The brand you fight against at trade shows may not be the one the AI cites. Sometimes the competitor cited by the AI is a smaller player but with a more carefully curated editorial presence. That’s your reference point, not the revenue leader.
Mistake two: copying the format without copying the level of depth. You see that the competitor cites data, and you put in data. But its data is taken from citable third-party sources, yours is made up to fill the paragraph. The model doesn’t tell the difference on the first pass, but the sources that would pick you up do.
Mistake three: ignoring the publications. The strongest pattern isn’t “what type of content” the competitor publishes, it’s “which publications it gets cited on”. Those publications are the highest leverage point for you: if you get into them, you get into the same pool the model draws from.
Mistake four: doing the analysis only once. AI answers change within weeks. A competitor analysis done in March and never repeated in September is worth nothing. It needs to be redone every two or three months on the same queries.
The operational audit to run in the next two weeks
Three concrete actions, in order.
- Choose eight queries that represent the need of your typical customer. Run them on ChatGPT and Perplexity. Identify the most-cited competitor.
- Extract the list of sources that the two engines cite for that competitor. Classify them by type and note on a sheet the three most recurring vertical publications.
- Check whether the competitor has a populated Wikidata entry and Organization schema on its homepage (Rich Results Test). Note what it has that you don’t.
Binary decision threshold: if the competitor has 3+ vertical industry publications citing it and you have zero, you’ve found the leverage point. If it has a Wikidata entry and you don’t, you’ve found the second. You don’t need any other metrics to get started.
From here on: the thread of visibility in AI answers
Reverse engineering the competitor is one of the five or six tools I use to help a brand enter AI answers in a controlled and measurable way. It’s not magic, it’s not a single factor: it’s one of the pieces that fits together with your editorial presence, with your entity authority, with the semantic structure of your pages.
In the following articles in this series I’ll explain how to track mentions over time, how to measure the sentiment of AI answers about your brand and how to build a monthly report that makes AI visibility a manageable KPI. All pieces that build on the work I’ve described here: without a map of the competitor, you don’t know where you’re aiming.
