The competitors the AI cites most in your industry have two or three specific sources in common that mention them — and almost never the outlets you would have guessed. Those sources are the missing link in your visibility, and until you know which ones they are you're looking for solutions in the wrong place. Reverse engineering their citations takes method, but the result is a short, concrete list of where to act.
The winning competitors share three sources you’re missing. They aren’t the obvious outlets, they aren’t the national newspapers, they aren’t the industry blogs everyone cites in their slides. Reverse engineering AI citations reveals your real editorial gap: the places where you need to be present in order to show up in the answers of Perplexity, ChatGPT and Gemini when a customer searches for your category.
Let me explain it with a case I just tested. Industry: growing cut flowers on the Ligurian Riviera, ranunculus and carnations. You’re a grower from Bordighera, you have a historic greenhouse, you sell to florists in Northern Italy and to wedding planners. You open Perplexity and ask “best ranunculus growers Ligurian Riviera”: the AI cites six companies, and you aren’t among them. The question isn’t “why doesn’t it cite me”. The question is: which sources is it pulling the names of the other six from?
What reverse engineering citations actually means
When Perplexity (and increasingly ChatGPT in search mode and Gemini) answers a query, it shows the links of the sources it consulted. This isn’t a cosmetic detail. It’s the explicit map of the information graph the model traversed to build the answer. If your competitor appears in the answer, it means that at least one of those cited sources talks about them in a way the model deemed useful for answering.
The principle is documented in the retrieval-augmented generation mechanisms I covered in earlier articles on tokenization and the weight of implicit citations: the model doesn’t “know” about companies, it pulls them from specific sources that are retrieved at query time. From this follows a direct operational consequence for you: the cited sources aren’t a byproduct of the answer, they are the cause of the answer. Working on those sources is the only direct lever you have.
I’ll warn you that here I’m proceeding empirically. I’m not citing a specific paper because competitive reverse engineering as an operational practice doesn’t (yet) have dedicated academic literature. Instead, I’m bringing you a test I ran with proper rigor.
Why this analysis comes before everything else
Before writing new content, before rebuilding the site, before working on recognizing your brand as an entity, you need to know where the models go fishing in your industry. Without this information you’re shooting in the dark: you produce excellent content on channels the model never consults for that category of query.
The sources the AI favors in a niche agricultural sector are different from the ones it favors in fintech or in the restaurant business. You can’t deduce them from your desk. You have to look at them.
Confusing “generally authoritative sources” with “sources the model consults for that query”.
The test you can run in twenty minutes
Open Perplexity (the free version is enough) and prepare a list of 8-12 queries that a real customer would make in your industry. For the Bordighera grower they’d be things like:
- “best ranunculus growers Liguria”
- “carnation suppliers Riviera for florists”
- “cut flower nurseries Imperia”
- “wedding flower growers Liguria”
- “historic floriculture companies Ligurian Riviera”
For each query, do this:
- Run the query, read the answer
- Note down the cited competitors (company name)
- Open Perplexity’s sources panel and note down the URLs that generated those citations
- Record everything in an Excel sheet with three columns: query, cited competitor, source URL
After 8-12 queries you have a matrix of 40-80 rows. Now sort by source domain and count the occurrences. The sources that appear 3+ times are your editorial gap.
Identify the 3-5 most recurring sources in your competitors’ citations.
The test I ran myself
I took the real case of the Bordighera grower (a floriculture SME with five hectares of greenhouses, 40 years in business, B2B sales to florists and wedding planners). I mapped six competitors cited by Perplexity across the western and eastern Ligurian Riviera. For each one I ran 10 commercial queries and noted down the sources.
Result over the sample of 60 answers (indicative test, not a statistical study):
- Three sources recurred in over 65% of the citations of the six competitors: the portal of the Sanremo Flower Market, an Italian floriculture industry magazine (with an article archive accessible online), and a regional directory of Ligurian agritourism.
- General-interest outlets like Repubblica or La Stampa appeared in 12% of cases, but only for specific local news articles from 2019-2022.
- Wikipedia appeared in 4 answers out of 60, always for territorial entries (not company ones).
The Bordighera grower I analyzed was absent from all three recurring sources. Translated into practice: he’s been growing flowers for 40 years, sells to satisfied customers, has a decent website, but to the AI model he doesn’t exist because he doesn’t exist in the places the model consults. Honest limitation of the test: a sample from a single industry and a single macro-geographic area, but the pattern is clear enough to act on.
The mistakes I see most often
Confusing “generally authoritative sources” with “sources the model consults for that query”. An entrepreneur told me: “I have an article in Sole 24 Ore, so I’m authoritative”. True in an editorial sense, irrelevant if for his category of query the model pulls from three niche industry directories that it ignores.
Reverse engineering on a single query. One query is noise. You need 8-12 similar queries to see which sources recur. Recurrence is the signal, a single appearance is coincidence.
Looking only at the big competitors. The model often cites small players that are well positioned on specific sources too. Those small players are the most informative: you understand exactly what they did that you didn’t.
Limiting yourself to Perplexity. Perplexity shows the sources, ChatGPT and Gemini in search mode do too, but with different interfaces. Replicate on at least two engines to avoid optimizing for a single retrieval system.
What to do concretely next week
- Extract 10 commercial queries from your industry (ask your sales team which questions cold customers ask).
- Run each one on Perplexity, save answers and sources in a sheet.
- Identify the 3-5 most recurring sources in your competitors’ citations.
- For each source, check: can I be there with a company profile? Can I publish content? Can I get an editorial mention? How much does it cost, how long does it take?
- Compare with the 3-5 competitors the AI cites in your industry: what do they have on those sources that you don’t? A complete profile? A dedicated article? A listing with structured data?
Once you’re present, also verify that your company name is read as a single entity: recognizing the brand as an author entity (Author Entity) determines whether that citation actually gets associated with you or gets dispersed.
This is an entry-level analysis. It gives you the map of the gap, not the coverage strategy. For the publishing plan and monitoring over time you need professional tools and someone who cross-references the data with organic performance.
Where does all this lead?
Reverse engineering citations is the starting point for any serious strategy of visibility in AI answers. Without the map of your industry’s sources you’re producing content blindly. With the map, you know exactly where to invest the next hours of editorial work to be present when the model builds the answer that concerns you.
In the next articles in this series on how to measure AI visibility, I’ll explain how to turn this map into a structured citation building plan, how to monitor how often your brand enters and exits Perplexity’s answers over time, and how to distinguish the citations that convert from the purely decorative ones.
