Traditional keyword tools tell you what people search on Google. GEOly now shows you what AI searches for on behalf of your category — and, behind every query, who shows up and where you're missing.
Key Takeaways
- Before it answers, AI runs its own burst of searches (query fan-out). What really decides whether you get recommended is those dozen-plus retrievals AI runs behind the scenes — not the single sentence the user typed.
- GEOly's new AI Search Keywords captures every retrieval AI actually runs in your category, clusters them into demand roots, and shows how often AI searched each one, whether you show up, who does, and which pages AI read.
- Take the luxury jeweler Tiffany: AI ran 7,322 retrievals across 283 topics, clustered into 467 demand roots — and Tiffany appeared in only 62 of them. On the other 405 topics AI cares about, it's invisible.
- 82% of raw fan-out queries appear only once; clustered into roots, the recurrence rate climbs above 98% — which is the first time "presence rate" becomes a statistically meaningful, continuous metric.
- One signal worth watching: in Tiffany's category, 11% of AI retrievals explicitly included "reddit" — AI increasingly goes to communities for real, unvarnished opinions.
The overlooked truth: before AI answers, it searches first
When a shopper asks ChatGPT, Perplexity or Google AI Mode "what are some good jewelry sets" or "what wireless earbuds should I buy for $500," AI doesn't answer straight from memory.
It first fires off a burst of searches — breaking that one sentence into a dozen-plus more specific queries (the industry calls this query fan-out), pulling pages from across the web, digging through Reddit, reading reviews — then blends that raw material into the recommendation it hands the user.
What really decides whether you get recommended is not the sentence the user typed, but the dozen-plus searches AI typed behind the scenes. The problem: until now, those searches were completely invisible to you.
- Traditional SEO keywords = what people type into a search box.
- AI search keywords = what the machine types while answering.
They are not the same thing. You can optimize for the first for a decade and still be entirely absent on the second.
Introducing AI Search Keywords
Starting today, GEOly captures every retrieval AI actually runs in your category, clusters them into demand roots, and tells you: how many times AI searched each one, whether you show up, who does, and which pages AI read.
Start with the big picture. Open the brand page for luxury jeweler Tiffany and the top of it is an "AI health panel": AI visibility of just 2.9, coverage across 283 topics, an estimated 8.8K AI-driven visits per month, a corresponding $49.8K in estimated monthly revenue, and 4.1K cumulative AI mentions. Below that, "How AI sees Tiffany" lays out its strengths (Style, Craftsmanship, Design) and weak spots (Price −165, Value −59) one by one, while "Category performance" marks its AI rank in each sub-category (Hand Mirrors #2, Jewelry #3, Engagement Rings #6, and so on).
These are all outcomes — a century-old luxury house with single-digit AI visibility. The new feature this article is about, AI Search Keywords, answers a more important question: why that outcome, and where to start fixing it.
Scroll down and you reach the main event. The first line is a one-sentence checkup of the whole category:
AI ran 7,322 retrievals across 283 topics, clustered into 467 demand roots. You appear in 62 / 467 roots · 11% of retrievals point to Reddit · Gaps worth noting: 0 confirmed + 69 early signals.
In plain terms: across the broad jewelry category, AI searched 7,322 times to answer users; those searches cluster into 467 demand topics, and Tiffany — a century-old luxury house — appeared in only 62 of them. On the remaining 405 topics AI is actively looking at, Tiffany is invisible.
This is keyword research for the AI era: you stop guessing what people will search and look directly at what AI is searching.
Two definitions worth spelling out
First: why "roots" instead of raw search terms?
Look at one real record. When AI answered "Which women's bracelets have the best craftsmanship and finishing?", it ran a live search, and the raw fan-out query was best craftsmanship women's bracelets brands hand finished — the atomic term AI actually typed while answering. Open any answer record in GEOly and you can see that raw query, plus which brands that answer mentioned (13 in this one).
The catch: these raw terms are improvised fresh on nearly every answer and rarely repeat — in our data, 82% of raw fan-out queries appear only once, so on their own they have none of the stability of "keywords." So GEOly aggregates tens of thousands of these raw terms into "roots" by their 2–4-word phrases, pushing the recurrence rate above 98% — which is the first time presence rate becomes a statistically meaningful, continuous metric. The atomic terms become detail inside a root's drill-down; the unit of analysis in the product is the root.
Second: what does "present" actually mean?
Present = in the AI answer that contained this term, your brand was identified (same definition as our visibility / Share of Model leaderboards, including citations, shopping and other visible positions). It does not mean "this query's search results returned you directly" — it measures whether AI, while handling this kind of demand, ended up writing you into the answer. Also: retrieval count is the number of answer records we monitor, not public search volume.
Four layers you couldn't see before
1. Root map: what AI is actually searching in your category
Sorted by retrieval volume, the demand roots AI searches most in your category are laid bare, each one marked with who shows up and your presence rate:
- engagement ring — 716 AI retrievals · your presence 6% · present: Brilliant Earth, Blue Nile, James Allen
- long term (attribute) — 532 AI retrievals · your presence 11% · present: Brilliant Earth, Tiffany, Blue Nile
- jewelry sets — 339 AI retrievals · your presence 11% · present: Tiffany, Mejuri, Brilliant Earth
- jewelry brands — 266 AI retrievals · your presence 30% · present: Tiffany, Mejuri, Cartier
- wedding band — 258 AI retrievals · your presence 12% · present: Brilliant Earth, Tiffany, Quince
- pearl jewelry — 248 AI retrievals · your presence 11% · present: Mikimoto, The Pearl Source, Tiffany
At a glance you can see which terms you're holding (jewelry brands, 30%) and which high-heat terms you're all but absent on (engagement ring, 716 retrievals, only 6% presence). Roots also distinguish category words from attribute words (like long term, waterproof), and you can filter by either in one click.
2. Priority actions: terms AI is actively searching but you're absent on
The system automatically pulls out the terms AI searches often but you don't show up on, drops them into "Priority actions," and tells you who was present at the time:
- buying guide — 81 retrievals · you present 1% · present: Brilliant Earth, Quince, Bliss Diamond
- jewelry trends — 35 retrievals · you present 0% · present: Ettika, Ana Luisa, BaubleBar
- necklace earrings — 125 retrievals · you present 10% · present: Giani Bernini, The Pearl Source, Macy's
This isn't a pile of terms to optimize — it's a list of "should have won, didn't." Every line is an AI recommendation slot a competitor took that should have been yours.
3. The actual search sentences: not guessed, the queries AI really typed
Open any root and you can see the raw search sentences AI actually typed, timestamps and all. Behind the "jewelry sets" root, for instance, AI really searched:
- "real pearl jewelry set buying guide"
- "best affordable jewelry sets look expensive"
- "luxury pearl jewelry sets solid gold platinum settings Mikimoto Tiffany"
- "best tarnish resistant cross jewelry sets reviews stainless steel gold vermeil"
- "highest rated fine jewelry sets long-term owner reviews"
These sentences are, in AI's eyes, the problem the user is really trying to solve. They're closer to real demand than any keyword planner — because this is what the machine itself computed it needed to search in order to serve the user. Under the same root, GEOly also lists the source topics it hit (affordable luxury / diamond / waterproof / cross jewelry sets, and so on), so you can trace your way to finer-grained demand.
4. The pages AI read during retrieval: where the "raw material" lives
Under the same root, GEOly also lists the pages AI opened while retrieving — the raw material fed into that answer:
- reddit.com — Best Engraved Bracelets 2025 · opened ×88
- macys.com — Jewelry Sets Wrapped & Ready · opened ×20
- whowhatwear.com — The Diamond Edit · opened ×18
- brides.com — Celtic Engagement Rings Guide · opened ×15
- vogue.com — The Story Behind Dua Lipa's Bulgari Jewelry · opened ×15
Where to build presence becomes obvious. To get into this root's AI answers, you either become a cited page yourself or show up in these high-frequency sources.
One aside: the "who's present" panel in a root's drill-down is a full leaderboard. The single "jewelry sets" root has 387 brands on the same stage — Tiffany (this brand) ×36, Mejuri ×29, Brilliant Earth ×28, Swarovski ×23, Cartier ×21, and so on. Your real standing on this track is laid out plainly.
One more layer: who AI named while retrieving
When AI runs a search, it often writes a brand name straight into the query (like "shokz openrun review"). GEOly ranks these brands that AI named on its own into a separate leaderboard — in Tiffany's category, the ones AI named most while retrieving are Sterling (740 times), Style & Co. (371 times) and Alexis Bittar (357 times).
This is a proxy for "retrieval-layer mind share": the brands AI reflexively reaches for even while just gathering material are the ones that truly own category mind share.
A signal you can't ignore: 11% of retrievals went straight to Reddit
In Tiffany's category, 11% of AI retrieval terms explicitly contained "reddit." On the root map, these terms get tagged "AI checks Reddit for opinions."
Behind this is a trend already underway: AI increasingly distrusts a brand's own marketing copy and instead goes to Reddit, forums and real reviews to find what users actually say. If your brand has no real, positive discussion in those communities, AI will struggle to write you into the answer on these roots.
A note: this measures AI's retrieval intent (where it chooses to look for opinions), not citations. Who ultimately gets cited is covered separately by GEOly's Citation module.
Every "gap" is split into "confirmed" and "early signal"
Terms with low presence get flagged as gaps, but GEOly doesn't paint with one brush — it grades statistical rigor, and the benchmark is "the presence you'd be expected to have in adjacent tracks":
- Statistically confirmed: at 95% confidence (computed by count of distinct questions, not by day-deduplicated records), your presence on this term is significantly below your overall level in surrounding tracks — meaning you have the standing but are clearly absent, and the conclusion is ready to act on.
- Early signal: presence is low and a pattern is emerging, but the sample isn't yet enough for a statistical verdict — useful for positioning early and watching. The data accrues every day.
"0 confirmed + 69 early signals" tells you at a glance which opportunities are still budding. Stake out the ones worth staking out, and wait for them to turn "confirmed" before going all in.
Why this is worth opening today
If earlier GEO tools told you whether AI mentioned you, AI Search Keywords tells you why AI didn't mention you — and which term to start winning back from.
- GEO / SEO / content teams: you get AI's real topic list — build content around what AI is searching, not what people guess. Every zero-presence, high-volume term is a clear content KPI, with the source to publish into (Reddit / trade media / your own site) already marked for you.
- Brand / marketing teams: on every demand root, see your gap versus competitors clearly, lock in the "should have won, didn't" terms, and spend your budget where you can flip the result.
- Growth / whitespace hunting: high-retrieval, zero-presence roots with no dominant brand yet are the richest whitespace of the AI era — first come, first served.
How to use it
Log in to app.geoly.ai → go to your brand page → open the AI Search Keywords section.
- Use "Root map" for the panorama; "View all roots" lets you filter freely by category / attribute and by retrieval volume / your presence, and expand adjacent categories.
- Open any root to see AI's real search sentences, present brands and source pages.
- Pick the first term to go after from "Priority actions" and start today.
AI's search box is open to you for the first time. Go take a look — what is AI actually searching for in your category?
About GEOly AI
GEOly is a GEO data platform for DTC brands — the brand-intelligence hub for the era of AI search and agentic commerce.
As ChatGPT, Google AI Mode, Perplexity, Gemini and Grok rewrite the rules of how brands get discovered, GEOly turns this "invisible AI battlefield" into quantifiable, optimizable daily metrics:
- Not just your own brand — check any brand and see the whole track's shape inside AI: visibility leaderboards, Share of Model, the AI shelf, ads and source citations.
- All pre-monitored, zero setup — works out of the box, covering the 6 major AI engines.
- From diagnosis (AI Search Keywords, visibility, source citations) to action (GEO Agent, content suggestions) to integration (MCP & API), it connects the full GEO loop.
In one line: if you care how your brand gets seen in the AI era, GEOly is the pair of eyes you should have.
Choose the plan that fits you
Whether you're a single-brand team just starting to watch AI visibility or a growth team covering multiple platforms and brands, GEOly has a tier for you. Save 10% on annual billing:
- Basic — $49 / mo: single-brand teams getting started · 1 brand / 1 seat · 30 monitored prompts · 1 AI platform (ChatGPT) · 900 responses / mo · 3,000 AI credits / mo · daily visibility monitoring, industry charts, email support.
- Grow — $149 / mo: scaling up · 1 brand / 2 seats · 50 prompts · 2 AI platforms (+Google AI Mode) · 3,000 responses / mo · 10,000 credits / mo · ChatGPT Ads analytics, MCP tools, live chat.
- Advanced (recommended) — $399 / mo: multi-platform + GEO Agent growth teams · 1 brand / 4 seats · 100 prompts · 4 AI platforms · 12,000 responses / mo · 30,000 credits / mo · GEO Agent, priority support, weekly insights.
- Plus — $999 / mo: multi-brand full solution · 5 brands / 10 seats · 500 prompts · all 6 AI platforms · 90,000 responses / mo · 100,000 credits / mo · agent automation, RESTful API, priority support.
Basic supports a free trial — sign up and you're in. Teams chasing multi-platform visibility and the GEO Agent should start from Advanced ($399 / mo). All four tiers include industry-level data access — see the GEOly pricing page.
Get started now
Register at www.geoly.ai and plug your brand into the first intelligence network of the AI era. Open AI Search Keywords today and see what AI is actually searching for in your category — and which recommendation slots that should be yours you're missing.
- Website and sign-up: www.geoly.ai
- Try it now: app.geoly.ai
- See plans / free trial: GEOly pricing page
For more, see the GEOly author page and the GEOly blog.
