GEO is where SEO was in 2003. The change is real, and it is being sold with far more certainty than anyone has. It also goes by AEO, LLMO and AI search optimisation, and the differences between those belong to the people selling them. Almost every AI-specific technique on offer is unmeasured. The things with evidence behind them are mostly good SEO and good engineering, which is harder to sell as new.
We only promise what there's evidence for. Most clients have heard the confident pitch already, and it rarely survives a second question.
What moves AI visibility, and how sure anyone is
| The measure | The evidence | What we do |
|---|---|---|
| Crawl access and server-rendered HTML | Established. Server logs and retrieval tests agree: assistants read the HTML the server sends, and none of the major AI crawlers ran JavaScript when it was last measured. | Server-render anything we want quoted, decided per page type before the build starts. |
| Ranking in ordinary search | Established. Assistants retrieve from search indexes, and Google says its AI features run on core ranking. Ahrefs found about three quarters of the pages cited in AI Overviews ranked in the top ten in 2025. Later studies report less overlap and disagree on how much, so ranking is a strong signal, and a weakening one. | Treat conventional SEO as the foundation of this work rather than a separate job. |
| Stable, addressable URLs | A precondition. Content with no address cannot be cited by anything. | Give every listing, filtered view and page of results a real URL. |
| Provenance on the page | A precondition. Answer engines extract the visible text and skip most metadata. | Author, date and version rendered in the copy, as well as in the tags. |
| Structured data (JSON-LD) | Mixed. Adding it produced no citation lift for pages already being cited, and assistants fetching a live page read the visible text rather than the markup. Microsoft says Copilot uses it. Google says its AI features need no special markup, while its rich results and its picture of an organisation do use it. | Emit the types with a documented use: Organization, Article, Product, Breadcrumb, Video. Claim what each is documented to do, never a citation lift, and put the same facts in the visible copy. |
| Rewriting copy for an engine | Weak. The original GEO study's gains came from rewording a page that had already been retrieved, in a lab. A 2025 benchmark found most of those tactics did nothing or hurt ranking, and a 2026 review of 45 studies found no technique with a lasting effect across platforms. | Write each claim and its evidence plainly, for readers. Copy rewritten for one engine can lose its place in the search that feeds all of them. |
| llms.txt | Weak. No provider has committed to reading it, and Google says it does not. Across 137,000 domains Ahrefs tracked, 97% of llms.txt files got no requests at all in May 2026. | Five minutes' work that nothing depends on. |
| Guaranteed citation or placement | Not achievable. No supplier can promise it. | Anyone who does is guessing. |
Structured content, clean rendering, real URLs and strong conventional SEO serve readers, assistive technology, search engines and answer engines at once. That is why GEO is not a separate workstream from accessibility, performance or search. A standalone "AI readiness" engagement usually turns out to be those three, relabelled.
The words and the reputation count as much as the build
Everything above gets a page read. Whether it gets quoted comes down to the words on it, and whether it gets found at all depends partly on who else is talking about it. Agentic SEO splits search into technical, on-page and off-page and says which of each agents do. The same split applies here, and this page is mostly the technical third.
On the page: an answer engine lifts a passage, not a page, so every section has to stand up on its own. The claim goes in the first sentence, nothing points back across a heading with a pronoun, and a specific beats an adjective every time. It is the same discipline good writing for people asks for. This site is written to a guide that asks for exactly that, for exactly this reason.
Off the page: in the largest study so far, the signal that tracked AI visibility most closely was not on the brand's own site at all. Across 75,000 brands, mentions elsewhere on the web correlated with appearing in Google's AI Overviews at 0.66, against 0.22 for backlinks. The authors are clear that this is correlation. A 2025 University of Toronto study points the same way: compared with Google, AI search leans heavily on third-party coverage over a brand's own pages. Coverage like that is earned by doing something worth mentioning, and no build decision gets you there.
One source, as many formats as anything needs
Content is never kept twice. Two copies of the same claim drift apart, and the drifted copy is the one a machine quotes. That doesn't rule out machine-facing formats: a sitemap, a feed and a markdown version can all be generated from the one source. Better still is serving the other format at the same URL, chosen by the request's Accept header, so there is only ever one address to keep right. Services that sit in front of a site and hand AI bots their own "AI-optimised" version are the second copy by another name.
Every page on this site works that way, and carries its provenance in both. The byline under each position reads the author's name and Made by Many, the markdown version opens with the same and the dates, and the structured data says the author works for the studio that publishes it.
What that means for a build
All of it comes down to decisions made before the first page is built. Retrofitting any of them costs more.
Readable without JavaScript
- Server-render editorial content. Anything that only appears after hydration doesn't exist for an AI crawler. Decide the rendering strategy per page type, up front.
- Put the figures in the markup. Every chart gets a table behind it, so the numbers exist as text and not only inside an image or a canvas.
- Move claims out of PDFs and onto pages. A figure in a PDF gets scraped or half-remembered. A figure at a URL can be cited, and corrected.
Addressable down to the claim
- Give everything an address. Listings, filters and paginated results need real URLs, with the filter and page number in the path or query. The usual failure is a listing that loads results client-side and never changes the URL, so nothing past the first screen can be reached, linked or cited.
- Anchor sections and tables, so the thing cited is the claim rather than the whole document.
- Return real status codes. Around a third of assistant fetches land on 404s, against eight per cent for Googlebot. An app shell that answers 200 to everything teaches a crawler that every wrong guess is a real page.
Attributed, and cheap to fetch
- Model provenance as structured fields, and render them. Authorship, date, version, revision status and how a page relates to others, held in the content model so they are consistent without relying on every author, and shown in the text the assistants actually read.
- Build the machine-facing surfaces in. Sitemaps segmented and regenerated on publish, feeds for anything time-ordered, and where the content warrants it a documented data endpoint with filtering and paging in the URL.
- Set cache headers per page type. Framework defaults often stop every downstream cache holding a page, which quietly turns every request into an origin request. Check this on any inherited estate.
Uncitable content is expensive content
A client worried about the cost of machine traffic and a client worried about AI visibility are usually looking at the same defect. A listing rendered in the browser with no URL per page forces anything that wants the data to crawl the interface, the most expensive route there is, and nothing it finds can be cited back. Give it addresses and both problems go.
Scores and scanners
Agent-readiness scoring arrived in 2026. Run it, and claim nothing for it, because nothing yet connects a score to an outcome.
Where AI visibility data comes from
Nobody can tell a client how often an assistant cites them, or why, because the providers mostly don't report it. The nearest thing to a reason is Microsoft's grounding queries, which show what a page was retrieved for, though not why it was chosen over the others. Some first-party data is free and worth collecting.
| The source | What it shows | Worth knowing |
|---|---|---|
| Bing Webmaster Tools | Citations in Copilot and Bing's AI answers, page by page, with the grounding queries behind them and each page's share of the citations for a query. | The only query-level AI data any provider publishes. Grounding queries are the short searches the AI wrote for itself, not what people typed, and the counts are samples rather than a full record. There is no API yet, so export it monthly. Bing's index also feeds ChatGPT search, so it is a partial proxy there. Verify every client site; it takes minutes. |
| Microsoft Clarity Citations | The same kind of citation and grounding-query data, beside Clarity's behaviour data, with AI referrals as a share of sessions and a branded and non-branded split. | Free on any site already running Clarity. One Clarity project per domain, verified through Bing Webmaster Tools or Search Console. |
| Google Search Console | AI Overviews and AI Mode impressions, in a generative AI report of their own since June 2026, by page, country and device. | Impressions only: no queries, no clicks, and no split between AI Overviews and AI Mode that we could find. A setting can now exclude a site from Google's AI features, so check it is set to include. |
| Analytics referrals | Visits from assistants, in GA4's AI Assistant channel or an Adobe segment. | A floor. Plenty of assistant traffic arrives with no referrer and lands in Direct. |
| Server logs and bot traffic | Which pages ChatGPT, Claude and Perplexity fetch live while answering someone, kept apart from the crawlers that collect training data. | The most checkable sign that a page is used in answers, and it is the site's own data. It says nothing about what was asked or whether the page was cited. Vercel added an AI Traffic view in October 2026; per-bot queries need its Observability Plus add-on. |
| OpenAI, Anthropic, Perplexity | Nothing for site owners. | Crawler documentation only. Each runs separate bots for search, for training and for fetches made on a user's behalf, and blocking the search bot can keep a site out of their answers. |
The commercial tools build both halves of their dataset themselves: prompts run at scale against the assistants, and prompt volumes estimated from bought clickstream panels, mostly browser extensions. It is the same method, with the same error bars, as SEO keyword volumes, so read the numbers as estimates. Profound is the enterprise option, and its headless browser sessions sit closer to what users see than API calls do. Semrush's AI Visibility Index makes sense where a client already pays for Semrush. Otterly and Peec are the lighter end, and AthenaHQ has a free tier good enough for a trial.
The answers themselves barely repeat. In SparkToro's study of almost 3,000 runs, made with Gumshoe, which sells tracking tools, the same list of brands came back less than 1% of the time, and in the same order less than 0.1%. A rank or position from any of these tools is noise. How often a brand is mentioned across many runs of the same question is the figure that holds up, and it is the one we report.
None of them checks whether the answer is right. They measure share of voice, because their buyers are marketers. Where a client's content carries authority, accuracy and currency matter more than presence, and nothing on the market covers that. Testing it takes the same synthetic-prompt method the paid tools use, pointed at a set of questions with known correct answers, which is how we run it (below).
One caveat applies to all of it. The API with search and the consumer apps are different surfaces, so anything measured this way shows how the systems retrieve rather than exactly what a given person sees. Good for trends, not for any one user.
How we run it for a client
On a new build, the three steps above are part of building it. mxa-discoverability puts in the rendering decisions, the addresses, the provenance and the machine-facing surfaces as each page is made, and the checks on every pull request keep them there.
On an existing site, we start with mxa-audit-discoverability. It samples three pages per template from the sitemap, scores crawlability and metadata with Lighthouse, runs a request-level sweep for whether an AI crawler can read the page at all and whether its metadata survives the scripts, validates the structured data, and reads each template for citability by eye. It produces two reports: a summary for whoever owns the outcome, and a fix spec for whoever does the work. Then it works the spec and confirms each fix landed. Every report also names what a fuller audit would add beyond its sample: the engines' own data, a whole-site crawl, bot logs and a checked set of AI answers. The access we ask for on day one is listed in agentic SEO.
For monitoring, the free first-party sources above, set up properly: Bing Webmaster Tools and Search Console verified for every property, Clarity Citations switched on wherever the site runs Clarity, and the analytics channel for AI assistants switched on. Once a month we read Bing's grounding queries and the assistants' fetches in the server logs.
For accuracy, which no tool on the market checks, a question set. Thirty to fifty questions, agreed with the client before anything runs, across four kinds: what the company is, where it appears in its category, how it compares with named competitors, and plain facts with a known correct answer written down first. Each is asked of ChatGPT, Claude, Perplexity, Gemini, Copilot and Google's AI features at least three times, and a person scores whether the client is mentioned, which sources were cited, and whether what was said is true. The result is a range and a list of recurring sources, never a single share-of-voice figure. A fact the client's own site doesn't state anywhere is a finding before a single assistant is asked.
Sources
- Google Search Central: AI optimisation guide, May 2026. Google's own statement that its generative features run on core Search ranking.
- Ahrefs: schema markup and AI citations, May 2026. The controlled study behind the structured data row: 1,885 pages that added JSON-LD against 4,000 controls, limited to pages already cited in AI Overviews, with every schema type pooled.
- searchVIU: schema markup and AI, October 2025. A price placed only in JSON-LD was found by none of five assistants fetching the live page.
- Search Engine Land: Microsoft Bing and Copilot use schema, March 2025. Fabrice Canel of Bing, on stage at SMX Munich. A vendor statement, not a measurement.
- Ahrefs: AI Overview brand visibility factors, May 2025. The 75,000-brand correlation study behind the off-site paragraph.
- searchVIU: AI crawlers and JavaScript rendering, December 2025. Retrieval experiments behind the server-rendering row.
- Vercel and MERJ: the rise of the AI crawler, December 2024. Server-log analysis behind the crawl-access and status-code figures.
- Ahrefs: search rankings and AI citations, July 2025. The study of a million AI Overviews behind the three-quarters figure in the ranking row.
- Bing: AI Performance in Bing Webmaster Tools, February 2026. Citation counts, cited pages and grounding queries, and Microsoft's note that the figures are samples.
- Bing: Intents, Topics, Citation Share and Compare, June 2026. The share of citations a page takes for a grounding query.
- Microsoft Clarity: Citations now generally available, May 2026. Grounding queries and cited pages inside Clarity.
- Google Search Central: generative AI performance reports, June 2026. Search Console's report of AI feature impressions, by page.
- Vercel: AI traffic insights in Observability, October 2026. Assistant retrievals and crawler traffic, by bot and page.
- SparkToro: AIs are highly inconsistent when recommending brands, January 2026. The repeat-run study behind reporting mention frequency rather than position.
- Aggarwal et al.: GEO, Generative Engine Optimization, 2023. The original study: gains from rewording one source among results already retrieved.
- Puerto et al.: C-SEO Bench, 2025. Most conversational-SEO tactics were ineffective or harmful, and conventional SEO did better.
- Chen et al.: Generative Engine Optimization, how to dominate AI search, University of Toronto, 2025. AI search favours earned, third-party sources over brand-owned content.
- Martinez: a critical survey of generative engine optimization, July 2026. A review of 45 studies; no technique shows a stable effect across platforms. A preprint.
- Ahrefs: llms.txt study, June 2026. Requests to llms.txt files across 137,000 domains.