Twelve weeks. 135,700 AI citations. One real website. I tested what actually makes ChatGPT, Perplexity, and Google's AI cite you — and most of what the new “GEO experts” are selling didn't survive contact with the data.
In the last twelve months, every SEO consultant on earth became a “GEO expert.” Same playbook, new acronym, a higher invoice. Most of them are guessing.
Their method is what I call correlation theater: publish a page, check whether ChatGPT mentions you a week later, declare the tactic a winner, sell the course. No baseline. No controls. One screenshot dressed up as proof.
I went the other way. Over twelve weeks I ran 87 controlled experiments across 49 pages of a live site — GrantCompass, a Canadian funding-discovery tool I built and grew to tens of thousands of users — and measured every change against real citation data. Microsoft's Copilot alone cited those pages 135,700 times between January and June. That's the dataset talking in this piece, not a vendor's pitch deck.
Here's the uncomfortable part: most of the advice you've read about getting cited by AI is wrong. Some of it I believed myself — until my own data killed it. Let me show you.
“GEO is just SEO with a new name. Your rankings will carry over.”
they're now two different games, played on two different fields.The overlap between Google's top-10 results and the pages AI engines actually cite has collapsed from roughly 76% to 38% in a single year. About 80% of AI citations don't rank in Google's top 100 at all. You can be invisible on Google and everywhere in ChatGPT — or the exact reverse.
Even GEO's advocates concede the split. As SEO strategist Aleyda Solis puts it, there's real overlap between optimizing for search and for AI — "although there are also important differences."
“Add schema markup and the AI engines will cite you.”
The most-repeated tip in GEO, and the cleanest failure in the data. A causal study of 1,885 pages found schema markup produced no citation lift whatsoever. AI engines read your visible words; they largely ignore the JSON-LD scaffolding underneath.
I'm not above this one. Schema was a Tier-1 tactic in my own playbook until the data forced me to demote it to “hygiene.” Keep it for the basics — it won't hurt. Just stop expecting it to win you anything.
“Write long, keyword-stuffed pages, and engineer every sentence to be quotable.”
structure wins, not length or clever sentences.All three lose. 53% of the pages cited by Google's AI are under 1,000 words. Keyword-stuffed URLs get fewer citations than clean ones. And the four most “GEO-coded” tactics I built — the ones that felt clever — all underperformed and got retired:
Hand-crafting definitional, comparison, and “quantified” sentences designed to be lifted verbatim. Result: 0 of 4 wins; one page dropped 79%. The engines pull well-structured passages, not individually-engineered lines.
Stamping every fact with “[Verified: date].” Result: median −60%. LLMs don't appear to reward per-fact date stamps — they read freshness at the page level.
Pre-writing the “question you might ask next” after every answer. Result: the worst median in the entire set, −90%. It bloated pages without adding extractable substance.
Packing five-plus numbers into every 50-word passage. Result: mild underperformance. Forcing density reads as noise. The lesson across all four: write naturally, structure deliberately.
”Optimize the page once, and you're done.”
citations decay — GEO is a standing program, not a one-time project.AI citations aren't permanent. The median citation half-life is about 4.5 weeks, and 40–60% of the domains an engine cites rotate every month. The work that won you a citation in March won't hold it by June — even if your page never changes — because the models keep re-deciding who the best source is.
So the goal isn't to “rank” once and walk away. It's to keep feeding the machine — fresh data, new answers, updated pages. GEO is a continuity program, not a campaign — which, conveniently, is also why you have to measure it continuously.
“Win one AI engine and you've won them all.”
a tactic that wins on one engine can actively lose on another.One of my pages dropped 20% on Bing while climbing on ChatGPT. In my own testing, FAQ-formatted content helped Google's AI but hurt ChatGPT. Only about 11% of cited domains show up on both ChatGPT and Perplexity. There is no single source the models universally trust — so optimize per engine, or accept that you're optimizing for one and guessing at the rest.
| ChatGPT | Perplexity | Google AI | Bing | |
|---|---|---|---|---|
| Publish original data | ✓ | ✓ | ✓ | ✓ |
| FAQ-format content | ✗ | · | ✓ | ~ |
| Question-style headings | ✗ | · | ~ | · |
| Citation-bait sentences | ✗ | · | · | ✗ |
| Schema markup | ~ | ~ | ~ | ~ |
| Persona-addressed sections | · | · | · | ✓ |
Here's the finding that matters most if you're small. The single biggest predictor of whether a page gained citations wasn't any tactic — it was the page's starting position. My same proven changes added citations to thin, new pages and lost them on established ones. Academic research backs the pattern exactly:
Here's what the "it's all off-site PR now" crowd gets wrong: I drove 135,700 AI citations to one site with almost nothing but on-page work — no PR, no link-building, no agency. On-site GEO is real, compounding leverage. These six things are what moved it.
Lead every section with the answer in 40–60 words. The engines lift your opening passage far more than your conclusion.
My single biggest win. A page that published proprietary numbers nobody else had became a citation magnet.
“If you're a first-time applicant…” Persona-addressed sections had the best win rate of any tactic I tested.
Quick answer → full explanation → deep dive. Progressive disclosure was the highest-volume winner in the set.
If/then eligibility trees and plain verdict statements (“the best option is X”) get pulled straight into AI answers.
Off-site isn't the whole game, but it compounds what your pages earn: brand mentions across the web predict citations ~3× more than backlinks. Get talked about where the models read.
Want the step-by-step version of all six? How to show up in AI search — the operator's playbook →
A single “did it cite me?” check is wrong about one time in nine, because AI answers are stochastic — the sources shift on every run. The only honest metric is your share of answers: how often you show up across many runs. Type a brand or topic and watch it happen.
I'm a growth lead for Invoice Simple and Joist at EverCommerce, a 2× founder, and ex-Meta. Over 11 years I've scaled DTC, SaaS, and subscription businesses — and built the tools, like LumenGEO, that measure them. I'm talking to early-stage teams about growth leadership and fractional work.
The 87-experiment dataset, the 135,700 citations, and the platform-by-platform findings are first-party — from GrantCompass via Bing Webmaster Tools and GA4. Several external figures above come from a single vendor (Ahrefs); I treat them as directional, not gospel. The FAQ-format finding is from my own testing, not an external study.