Every time AI cites you, the text it produces is packed with information: how it describes you, with which adjectives, in what context, alongside which competitors. If you capture that citation, post it on social media and forget about it, you're throwing away a competitive analysis that no market research would ever give you. Reading AI citations as data turns every mention into a signal about how the market perceives you — and about where you can improve your positioning.
When AI names you in its answers, that text is data. Extract it, clean it, analyze it: linguistic patterns, adjectives used, context. It’s pure qualitative intelligence.
I say it bluntly because in my articles I often see entrepreneurs who get excited at the first citation from Perplexity or ChatGPT, take screenshots, post them on LinkedIn, and then don’t know what to do with them. That text, on the other hand, is the richest raw material you have for understanding how AI sees you and how it sees your competitors. Let me explain how to handle it.
The hidden data inside an AI answer
Take a cashmere company from Spoleto, say a small knitwear maker in the Valnerina that works with Italian yarn and produces made-to-measure for European boutiques. Open Perplexity and ask: “best Italian producers of artisanal cashmere”. You get an 8-10 line answer with 4-5 brands cited, each with a sentence of context.
That sentence of context is gold. It doesn’t just say whether the brand gets cited, but how: “known for hand craftsmanship”, “specialized in made-to-measure garments”, “renowned for the use of premium yarns”. These are the adjectives that AI extracted from web content (website, reviews, industry articles, directories) and synthesized to describe the brand.
Mention mining is the process of taking those sentences, deconstructing them and turning them into a spreadsheet that becomes the foundation of your visibility in AI answers strategy.
Why it’s real competitive intelligence
In this series I’ve already talked about metrics like share of voice in AI answers and citation accuracy rate. Those tell you how many times you appear and whether AI describes you correctly. Mention mining goes one level deeper: it tells you with what language you appear and in which narrative context AI places you relative to competitors.
In the world of research on generative ranking there isn’t yet a definitive paper that measures the value of mention mining as an intelligence technique. It follows that the mechanism has to be built by deduction: AI summarizes the web before answering, so the adjectives and contexts it uses are the synthetic mirror of how the web talks about you. If AI says “known for stockinette stitch knitting” and you don’t say it on your website, it means reviews, industry blogs or directories are saying it: third parties are describing you, and that’s a fact that changes your editorial strategy.
Translated into practice: mention mining is the cheapest way to reconstruct the narrative that the digital ecosystem has built around your brand without you knowing it.
“What Perplexity says about [my brand]” is interesting once, then it becomes narcissism.
The test you can run in 30 minutes
I’ll give you the procedure I use with clients, calibrated for those who don’t have a team of analysts. You need a Google Sheet and access to a conversational AI engine.
Step 1 — build 10 realistic queries for your industry. For the Spoleto knitwear maker: “best artisanal Italian cashmere”, “cashmere producers Umbria”, “Italian cashmere premium yarn”, “made-to-measure cashmere producers Italy”, and so on. Mix brand-agnostic queries (“best X”) with geographic queries (“X in Umbria”) and functional queries (“made-to-measure X”).
Step 2 — run each query on 3 AI engines: ChatGPT, Perplexity, Gemini. Claude too if you use it. Save the text of the answer in a column in the sheet.
Step 3 — for each answer extract 5 fields: brands cited (in order of appearance), mention position (1st, 2nd, 3rd), adjective/context sentence associated with each brand, sentiment (positive/neutral/critical), sources cited by AI (URLs if Perplexity shows them to you).
At the end of the round you have a matrix. 10 queries x 3 engines = 30 rows. The pattern emerges from there.
Source audit: take the URLs cited by Perplexity in the answers.
The test I ran: three Italian cashmere brands
I’ll say it upfront: it’s an indicative test, not a study. I took three Italian cashmere brands (one internationally known, one national premium, one niche artisanal from Umbria) and ran 12 queries on ChatGPT, Perplexity and Gemini in Italian. Small sample, clear pattern.
Results across 36 total answers:
- The international brand appeared in 32 out of 36 answers, always among the first 3 mentions, described with adjectives like “iconic”, “globally recognized”, “a benchmark of luxury”.
- The national premium brand appeared in 21 out of 36 answers, average position 4th-5th, with adjectives like “traditional quality”, “Italian yarns”, “in-house production”.
- The niche artisanal brand appeared in 4 out of 36 answers, always when the query contained “artisanal” or a geographic reference, described as “small operation”, “made-to-measure craftsmanship”.
The interesting data point isn’t the frequency: it’s the vocabulary. The international brand captured the adjectives of luxury. The premium brand those of tradition. The artisanal brand those of the niche. Three distinct narrative territories, and AI is replicating them by inertia because that’s how Wikipedia, fashion magazines, marketplaces and industry blogs describe them. Anyone who wants to shift territory (the artisanal brand wanting to steal space from the premium one, for example) must first change the narrative that the ecosystem builds around them, and then hope that AI notices.
The mistakes I see most often
When clients try to do mention mining on their own, they always stumble on the same four points.
Querying AI only with your own brand name. “What Perplexity says about [my brand]” is interesting once, then it becomes narcissism. The useful query is the one without the brand, where AI has to choose who to cite.
Limiting yourself to a single engine. ChatGPT, Perplexity, Gemini and Claude draw from different sources and use different weights on brand authority signals. Measuring on just one gives you a third of the picture.
Ignoring the cited sources. When Perplexity shows you the 6 sources it extracted the answer from, those are your list of sites you need to be present on or buy future visibility from. It’s the map of the territory.
Confusing absence with defeat. If you don’t get cited, it doesn’t mean AI hates you: it means your brand doesn’t yet have enough structured signals in the right places — an entity on Wikidata, the knowledge graph, implicit citations on industry sites. It’s a problem of semantic infrastructure, not of likability.
What to do with the spreadsheet
Once you have the matrix, you do four practical things with it:
- Lexical map: extract the most recurring adjectives for each competitor. Compare them with those associated with your brand. If yours are poorer or more generic, the problem is positioning, not SEO.
- Source audit: take the URLs cited by Perplexity in the answers. The ones where competitors appear and you don’t are the sites where you need to work on editorial presence, partnerships or digital PR.
- Context gap: look for the queries where competitors appear but you don’t. Those define the narrative territories you’re missing. The editorial plan for the next 6 months starts from there.
- Temporal tracking: redo the same round of queries every 30 days. Save the answers. Compare the previous month’s adjectives with the current month’s. If they change in your favor, something in your editorial strategy is working; if they don’t change, you’re investing in the wrong places.
A note on author entity recognition: if in your industry AI cites people (designers, founders, yarn masters), mention mining should be extended to people’s names too, not just brands. For the Spoleto knitwear maker, knowing that AI associates the founder’s name with “Umbrian textile school” or “supply chain tradition” changes the owner’s personal branding plan.
It has to be said honestly: this is an entry-level check. Professionally structured mention mining requires dedicated tools that monitor hundreds of queries continuously, normalize the vocabulary with NLP and produce temporal dashboards. What I’ve described here is the manual work that lets you understand whether the game is worth the candle before investing in tools.
Mention mining is the fact-checking of your positioning
In this series I’m explaining how to measure visibility in AI answers seriously. Mention mining is the qualitative piece: it doesn’t tell you how visible you are, it tells you how you are visible. It’s the difference between knowing you have 100 mentions and knowing that 80 of those call you a “small artisanal operation” while your competitor gets called “a benchmark of the sector”. Same volume, two different positionings, two different commercial futures.
In the next articles of the series I go into detail on how to build a monthly AI visibility scorecard, how to deconstruct competitors’ content strategy and how to do gap analysis by query families. Mention mining is the first building block: without it, the other metrics stay abstract.
