Run the same prompt through ChatGPT seven times in an hour and you’ll get, on average, seven different lists of brands. Ask again tomorrow and the brand sitting at #1 today might not appear at all.
That’s not a flaw in your tracking tool. It’s the finding of a study Rand Fishkin and Patrick O’Donnell published in January 2026: 600 volunteers ran 12 prompts through ChatGPT, Claude and Google’s AI nearly 3,000 times, and the odds of two runs returning the same brand list came out under 1 in 100. The odds of the same list in the same order were closer to 1 in 1,000.
I’ve read a lot of “how to rank #1 in ChatGPT” content this year. None of it mentions that number.
Most GEO (“generative engine optimisation” — getting your brand mentioned or cited in AI-generated answers, as opposed to ranked in a list of blue links) advice published this year skips that fact entirely. It tells you to add schema, write in Q&A format and update your freshness date — all correct, all standard practice we run on every account before anything below — then jumps straight to “how to rank #1,” a sentence that assumes a #1 exists to hold onto.
What follows is the rest of our SOP: 24 tactics we run once the basics are done — the ones we covered in our earlier piece on the fundamentals of getting cited in AI answer engines — grouped the way we group them internally: measurement, training-data presence, page architecture, entity gaps, and platform-specific plays. Two more line items live in the same document, filed under R&D only. We’ll tell you why they’re not here.
Key takeaways
- A single check of your brand’s AI visibility is closer to a coin flip than a metric — the odds of an AI tool returning the same brand list twice are under 1 in 100 (SparkToro & Gumshoe.ai, January 2026).
- Mentions, citations and sources are three different things inside an AI answer. Most GEO advice treats them as one.
- Reddit alone accounted for 16.7% of ChatGPT’s tracked citations as of July 2026 (Ahrefs Brand Radar) — a genuine UGC presence isn’t optional any more.
- The correlation between classic Google rankings and AI-answer visibility sits at roughly 0.65, per Robert Niechciał’s research — high enough that SEO still matters, low enough that it won’t carry you alone.
- Models answering “what is X” pull from a more stable layer than models answering “which is best” — claim the definition before you chase the ranking.
- Two tactics in our own SOP — minting a brand-new entity outright, and serving AI crawlers different content than human visitors — stay in R&D only. Neither belongs in client work.
- None of this replaces classic SEO. It sits on top of it.
Measurement: why one check is worse than none
Everything below assumes you can tell whether a tactic worked. Most GEO reporting can’t, because it measures the wrong thing, at the wrong frequency, in the wrong place.
1. Track presence longitudinally, not once
Measure your brand’s presence in AI answers at least four times a month — roughly weekly — across a set of 100 or more prompts, and report the percentage of answers you appear in, plus how often you rotate in and out of the top results. Not your rank on one given day. Reject any report built from a single check; per the SparkToro and Gumshoe.ai study cited above, a lone snapshot tells you almost nothing about where you actually stand.
2. Scrape the public interface, not the API
Pull your tracking data from the public, logged-out versions of ChatGPT, Gemini, Perplexity and Copilot — the versions an actual customer would use — not their APIs. The API often runs a different system prompt and a different model tier, so it answers differently to a human typing the same question. Claude has no equivalent public, scrapeable surface yet, so leave it out of this exercise for now.
3. Query from simulated personas, not a blank slate
Build two or three persona profiles — a plausible history and professional context each — and run your prompts from inside them, then compare the results against an anonymous, no-history query. These models personalise answers per user profile, so the “neutral” version you get logged out isn’t necessarily what your actual buyer sees.
4. Diagnose your mention-to-citation ratio
For every brand you track, including your own, work out the ratio between how often it’s mentioned by name and how often it’s actually cited with a link back to its own domain. High mentions with low citation means you need to defend the visibility you already have. Low mentions with high citation means you should push harder for citations specifically. We’ve reviewed a Senuto case study of one household brand where mentions ran roughly three to one against citations — plenty of visibility, almost none of it pointing back to the brand’s own site. (We’re keeping the brand anonymous here pending sign-off — happy to name it if you want the specificity.)
5. Monitor across six intents, not one
Split your prompt set into six intent types: definition, validation, exploration, comparison, ranking and recommendation. Measure each separately, and you need fewer prompts per intent than you’d need for one undifferentiated set. In our own tracking, the definition intent holds steady in the high 50s (percent of stable answers), while recommendation swings as low as the low 30s — visibility in one intent tells you almost nothing about the others.
6. Test for training-data presence with web search switched off
Turn off web search and ask the model to recommend a brand in your category. If yours comes up unprompted, with no search running, that’s a sign you’re already baked into the model’s training corpus. Mark that intent “secured” and put your effort somewhere that isn’t already won.
Training data: getting into the next model, not just this search
Everything above measures the search-and-retrieval layer. This section is about the layer underneath it: the corpus these models were actually trained on.
7. Publish a genuine English version of your key pages
If your primary market isn’t English-speaking, publish a full, editorially considered English version of your most important pages — not a machine translation bolted on for appearances. Common Crawl, the open web archive most large language models train on, is overwhelmingly English-language. A thin auto-translated page won’t earn the same weight in training as one actually written to be read.
8. Check robots.txt and let CCBot in
Confirm you’re not blocking Common Crawl’s crawler, CCBot, in robots.txt, and actively link to and ping your most important URLs. Common Crawl’s archive has grown past 300 billion pages, and it’s a primary feedstock for the next generation of model training. The target here isn’t this month’s search results — it’s next year’s model.
Page architecture: build for extraction, not persuasion
Robert Niechciał, CTO of Vestigio, said it to a room of a few thousand SEOs on a webinar for his Sensei Academy platform: “AI cites fragments, not whole articles.” Structure follows from that one sentence.
9. Let your title and meta answer the question outright
Write your meta description as a genuine summary — the primary entity plus the core fact — specific enough that a model could answer the query from the snippet alone, without ever opening the page. Several AI systems answer directly from a search snippet more often than most SEOs assume. A meta description written as clickbait works against you here, not for you.
10. Mark up entities and relationships in JSON-LD, then test it
Add schema with explicit entities (a model’s term for a distinct, named real-world thing — a company, a person, a product) tagged as Organization, Person or Product, plus sameAs links tying them to your other verified profiles. A/B test it across a set of subpages and measure the actual change in AI-answer presence. JSON-LD is the cheapest structure a model can extract; it skips parsing your HTML and JavaScript entirely. Don’t assume it works. Measure it against a control set of pages that don’t have it.
11. Write self-contained, question-and-answer paragraphs
Each paragraph should work as its own standalone unit: a short question, a short answer, the supporting data, then stop. “Every paragraph is a self-contained unit of knowledge,” as Niechciał put it in the same webinar — because whatever model eventually cites you is lifting a fragment, not the article. Use lists only where the content is genuinely list-shaped; in practice, most AI-cited content is prose, not bullets.
Entity gaps: publish exactly where the gap is
12. Build a gap matrix before you write anything
Cross-reference competitor, intent, category and funnel stage in a single matrix, and mark every cell where a competitor shows up and you don’t. Publish exactly into those cells. It’s reverse-engineering instead of guessing — you stop firing blind and start aiming at documented absence.
13. Claim the “definition” intent for your core category
For the category you most want to own, publish standalone content answering “what is X” — with your brand named as part of the definition itself, not appended afterwards. Definition-type queries sit on the most stable of the six intents from the measurement section above. Win it, and it tends to stay won.
14. Dominate the “X vs Y” comparisons
Build and seed comparison content — your own pages plus genuine user-generated posts — for every “X vs Y” and “X or Z” query in your space, aiming to be the default name mentioned in any comparison. In the same Senuto case study behind point 4, the bulk of that brand’s AI mentions came specifically from comparison-style queries. Comparisons carry outsized weight relative to how often people actually ask them.
15. Buy placements exactly where your competitor is already cited
Export the list of domains citing your competitor for a given intent, cross-reference it against a publication marketplace such as WhitePress, and buy placement on precisely those domains. It’s one targeted link instead of a hundred generic ones — the most surgical move on this list, and the one most agencies skip because it takes real research instead of a media-buying spreadsheet.
16. Repeat your brand across different contexts, not the same claim
Plan a content series where your brand shows up across genuinely different intents and contexts — a consistent story told from different angles — rather than the same thesis published a hundred times. Retrieval rewards a brand that turns up in varied contexts. It doesn’t reward repetition of a single claim.
Platform-specific tactics: treat each one differently
Optimising for “AI search” as a single channel is the same mistake as optimising for “social media” as a single channel. Each platform has its own retrieval habits.
17. Match your source strategy to the platform
| Platform | Leans toward | What that means for you |
|---|---|---|
| Gemini | Medical, government and academic sources | Earn or partner for citations on .gov, .edu and recognised health authorities where your category touches them |
| Perplexity | E-commerce sources | Product pages, marketplaces and comparison content carry more weight here than elsewhere |
| ChatGPT | User-generated content | Reddit, YouTube and category forums are the heaviest lever |
Cross-platform citation overlap is thin — in our own tracking, a source cited on one platform rarely shows up as a source on another. Optimising for one doesn’t carry over to the rest, so map your source targets platform by platform instead of treating this as one channel.
18. Build a real UGC layer
Get into Reddit threads, run a YouTube channel, show up in the forums your category actually uses. As of July 2026, Reddit alone accounted for 16.7% of ChatGPT’s tracked citations — more than any other single domain, per Ahrefs’ Brand Radar data. That’s not a growth hack. It’s where the model is already looking.
19. Publish full transcripts of every podcast and video
Every video or podcast episode should ship with a complete written transcript on your site. Models read text, not audio or video — the transcript is what actually gets ingested and cited, not the recording sitting next to it.
20. Show up on three or more platforms at once
Spread your effort so your brand appears across at least three AI platforms simultaneously, rather than concentrating everything on one. In our own tracking, brands with a genuine presence on three platforms capture a disproportionate share of total mentions — presence compounds across platforms rather than adding up in a straight line.
21. Fill in every map profile, not just Google
Complete your listings on Apple Maps, Bing Maps, OpenStreetMap, HERE, Waze and, where relevant, Targeo — photos, categories, description and links, to the same standard you’d use for your Google Business Profile. We picked this one up from a client engagement where map presence measurably fed AI-answer visibility in ways a Google-only listing didn’t.
22. Set up Bing Places, but don’t bet on IndexNow
Claim your Bing Places listing, but don’t treat Bing or the IndexNow protocol as your main lever into ChatGPT’s visibility. We’ve seen ChatGPT cite pages that don’t rank anywhere in Bing’s own results — the assumption that “visible in Bing” equals “visible in ChatGPT” doesn’t hold up in our testing, though we’d call this an informed pattern rather than a confirmed mechanism.
23. Target the top three in a narrow category, not visibility everywhere
Pick one narrow intent or category and concentrate your budget on reaching the top three there, rather than chasing “visibility” across everything you do. In our tracking, top-three presence holds up far better over time than a mid- or long-tail spot, which tends to turn over almost every time you measure it. Spreading thin buys you nothing worth reporting.
24. Choose your next category by saturation, not ambition
Before you invest, check how concentrated the answers already are in that category — how many brands make up 80% of the responses. Some categories are cemented around a dozen incumbents; others are wide open, with the same 80% share split across several hundred brands. Attack the fragmented ones. A new brand rarely cracks a cemented category through content alone, no matter how good the content is.
What we keep out of client work
Our internal SOP has two more line items, both filed under R&D only: minting a brand-new entity or category outright, and testing whether AI crawlers can be served different content than human visitors see. Both carry real legal and platform-policy exposure. Both stay on our own test properties. Neither goes near a client account without a lawyer in the room first — and if a vendor pitches you either one as a packaged “growth hack,” that’s the tell.
Stop doing these five things
| Stop doing this | Why it fails | Do this instead |
|---|---|---|
| Monitoring brand sentiment in AI answers | These models are built to recommend, not to criticise — negative sentiment is close to nonexistent in the answers we track | Redirect that budget to intent-based presence monitoring (point 5) |
| Publishing a one-off “AI visibility report” | A single check is close to a coin flip — see the SparkToro data above | Report longitudinally: minimum four checks a month, plus rotation (point 1) |
| Buying 100 identical brand mentions | Volume in a single intent and context doesn’t build coverage | Publish across different intents and contexts (point 16) |
| Running fake reviews | Inconsistency detection keeps improving; one penalised account can undo a month of work | Collect genuine reviews from real customers |
| Generating scale content on your most expensive model | Cheaper models match the quality bar for high-volume content at a fraction of the cost | Use cheaper models for scale, save premium models for final polish |
The bottom line
GEO isn’t new SEO wearing a costume, and you can’t report on it the way agencies reported rankings in 2015. The hardest fact to accept is also the most useful one: a measurement taken once is closer to a coin flip than a metric, so build the cadence before you build the content. What’s still genuinely unclear a year into this: whether the citation indexes behind ChatGPT, Gemini and Perplexity converge toward something like a shared web graph, or stay as fragmented as they are today. Until that settles, betting everything on one platform’s rules is the riskiest move on this entire list.
FAQ
How often should I measure my brand’s presence in AI answers?
At minimum four times a month, spread across a week, over 100 or more prompts. A single check tells you almost nothing — the odds of an AI tool returning the same brand list twice are under 1 in 100 (SparkToro & Gumshoe.ai, January 2026).
Does classic SEO ranking still matter if I want to show up in ChatGPT or AI Overviews?
Yes, but it’s not the whole story. Robert Niechciał puts the correlation between top-10 Google rankings and AI-answer visibility at roughly 0.65 — strong enough that SEO still matters, loose enough that ranking well won’t guarantee a citation on its own.
Should I track sentiment around my brand in AI answers?
Generally not worth a dedicated budget line. These models are built to recommend rather than critique, so negative sentiment is rare in the answers we track. Spend that effort on tracking whether you’re recommended, and in what context, instead.
Is buying a large volume of identical brand mentions a shortcut to AI visibility?
No. Volume in a single intent and a single context doesn’t build coverage across the six intents that actually decide whether a model recommends you. Spread across contexts instead (point 16).
What about cloaking content for AI crawlers, or inventing a new category to own outright?
Both sit in our SOP’s R&D column, tested only on our own properties, never on a client account, because both carry real legal and platform-policy risk. Treat any vendor pitching either one as a ready-made “service” with real suspicion.
If you want to know where your brand’s mention-to-citation ratio actually sits before you spend a single token on content, that’s what our AI Visibility Audit is for. Send us your domain and the platforms you care about — we’ll send back the gap matrix.
Rafał Moszkowcow, CEO & Co-Founder, NON.agency. Fifteen years running search operations, most recently building the AI-native infrastructure behind NON.agency’s international SEO work. Hosts the Globalne Horyzonty podcast.
