GEO in 2026: how to get cited by artificial intelligence
What actually works to get cited by ChatGPT and AI Overviews, and what to stop doing.
- Date
- July 21, 2026
- Category
- AI & Search
- Reading
- 8 min read
- Byline
- Outsmart Studio
Brand mentions correlate at 0.66-0.71 with visibility in AI answers. Backlinks, near zero. And llms.txt, chunking and AI-specific markup, the three things everyone is selling right now, Google states it does not use.
To get cited by ChatGPT, Google AI Overviews or Perplexity in 2026, what counts most is how much your brand is named elsewhere on the web, not how many backlinks you hold. Brand mentions show a 0.66-0.71 correlation with visibility in AI answers; backlinks, almost no impact at all (Ahrefs, 75,000 brands, December 2025). Meanwhile llms.txt, chunking and AI-specific markup are explicitly listed by Google as useless.
Why this became urgent now
Between January and April 2026, 68.01% of Google searches in the United States ended without a single click: it was 60.45% in 2024 (SparkToro on Similarweb clickstream data, June 2026). In Italy the figure is lower but still stark: 63.4%, with 280 clicks out to the open web per 1,000 searches.
The traffic has not vanished: it has changed shape. Google reports over 2.5 billion monthly users for AI Overviews and more than 1 billion for AI Mode one year after launch (Google I/O, 19 May 2026). In Switzerland 76.1% of adults use AI assistants, and 41.6% of those already using them search the web with them instead of Google or Bing: roughly one adult in three (Comparis/Innofact, March 2026, 1,035 people).
In plain terms: ranking first no longer guarantees a citation, and not ranking first no longer prevents one. Around 37% of cited pages do not appear in the top 100 at all.
What GEO actually is, without the marketing around it
GEO, short for Generative Engine Optimization, is the set of practices that raise the odds of a piece of content being retrieved and cited by a generative engine. The term comes from a paper presented at ACM SIGKDD in August 2024 by researchers from Princeton and IIT Delhi, built on a 10,000-query benchmark. The experiments themselves date back to the November 2023 preprint, which is not a minor detail.
It is not “SEO but new”. It is a different thing because the mechanism is different: the engine does not return a list, it breaks the question into many sub-questions, which Google documents as query fan-out, runs each one separately and then synthesises. Whoever covers the whole cluster of questions around a topic gets picked up more often than whoever optimises a single keyword.
- Mentions on YouTube
- Around 0.74. The single most correlated factor.
- Brand mentions across the web
- Between 0.66 and 0.71.
- Branded anchor text
- Between 0.51 and 0.63.
- Search volume on the brand name
- Between 0.35 and 0.47.
- Domain Rating
- Between 0.27 and 0.33.
- Number of backlinks
- Minimal impact.
These are correlations, not cause and effect: the authors say so themselves. But the direction matches another measurement, from May 2025: brands in the bottom half for web mentions get 0-3 mentions in AI Overviews; those in the top quartile average 169.
Generative engines read how much the world talks about you, not how much you talk about yourself.
- llms.txt
- “Google Search does not use it”. John Mueller compared it to the old keywords meta tag. A study of around 900 domains monitored for seven months recorded no requests to the file from GPTBot, ClaudeBot, PerplexityBot or Google-Extended.
- Content chunking
- There is no requirement to break text into tiny blocks.
- “AI-specific” writing
- No special language, no keyword-stuffed variant.
- Inauthentic mentions
- Scattering the brand name across the web does not work.
- Proprietary AI markup
- It does not exist.
On top of that comes the most counterintuitive finding of 2026: Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026, against roughly 4,000 matched control pages. The result: −4.6% citations in AI Overviews, +2.4% in AI Mode, +2.2% on ChatGPT. No statistically significant increase. And yet cited pages carry JSON-LD almost three times as often: schema is a marker of well-built sites, not a cause of citation.
- Foundations
- The page has to be indexed and snippet-eligible: watch out for nosnippet, max-snippet and data-nosnippet, which block the AI features too. Check that your CDN or WAF is not blocking AI crawlers.
- Brand signals
- Build real mentions off your own site, YouTube first. Reddit and YouTube are the only two sources in the top 5 across every platform; Wikipedia leads on ChatGPT; LinkedIn and G2 carry weight in B2B.
- Structure and writing
- Citing sources inside the content is worth around +27% visibility; improving flow and clarity, +28%; adding verifiable data, +30%. Keyword stuffing actively hurts, around −8%.
- Length
- Irrelevant. Across 174,048 cited pages, the correlation between length and citation is 0.04, and 53.4% of citations go to pages under 1,000 words.
- Language and domain
- In markets with strong regional brands, 46-59% of citations use national domains. For a company in Ticino that is a concrete argument for a .ch domain and content in Italian.
The structural ceiling nobody talks about
Anyone selling consulting or professional services should know this before signing a GEO contract: the ceiling is far lower than in tourism, where it reaches 23%. For Ticino that cuts both ways: hospitality has a lot to gain, professional firms much less.
A second under-reported figure: 58.8% of ChatGPT referral traffic lands on the homepage, not on the page that was cited. Per-page attribution for GEO is largely broken, and the homepage has to carry the conversion weight of visitors who arrive already informed.
How to measure it, and why to distrust tidy reports
AI Overviews change constantly: 45.5% of citations are replaced at every refresh, with an average persistence of 2.15 days. Semantic consistency, though, stays at 0.95 out of 1.0: the sources and the words change, the meaning does not. An “AI visibility score” measured once is noise.
- Check you are not being blocked
- AI crawlers, nosnippet, robots, WAF.
- Map the real questions
- The cluster of questions around what you sell, not the keywords.
- Rewrite the key pages
- Answer up front, data and sources cited.
- Move budget from links to mentions
- PR, YouTube, communities, real presence where your subject is discussed.
- Measure every week
- Not once a quarter.
- Make the homepage hold
- That is where people arriving from AI land.
GEO is not a new technical discipline: it is the old work of building recognition, finally with a way to track it.
- Does GEO replace SEO?
- No. 95% of ChatGPT users still use Google as well: AI adds to search, it does not replace it. And Google's AI features rest on its own ranking systems: no indexing, no citation.
- Should I publish an llms.txt file?
- It costs almost nothing and does no harm, but Google states it does not use it and no crawler from the major models appears to read it. It is not a visibility lever.
- Does schema markup help you get cited?
- Not demonstrably. A matched-control study across 1,885 pages found no significant increase on any platform. It remains useful for other reasons.
- How long should content be?
- There is no optimal length: more than half of all citations go to pages under 1,000 words. What counts is how clearly you answer.
- Is it worth updating old articles?
- On ChatGPT yes: it cites content on average 458 days more recent than organic results. On AI Overviews the effect is essentially nil.
Sources: SparkToro/Similarweb, June 2026 · Google, Guide to Optimizing for Generative AI Features, updated 10 July 2026 · Google I/O, 19 May 2026 · Ahrefs, December 2025, May 2025, March 2026, May 2026, November 2025, July 2025 · Aggarwal et al., GEO: Generative Engine Optimization, ACM SIGKDD, August 2024 · Similarweb, AI Search Stats 2026 · Comparis/Innofact, April 2026 · Profound, Index Report Summer 2026 · Peec AI via Search Engine Land, 31 March 2026.
