Field Manual · 15 min read

The AI Search Visibility Field Manual (2026)

Most GEO advice in 2026 still recommends llms.txt and schema markup first. The evidence does not support either as a citation lever. This is the field manual for the six levers that actually move AI citation, checked against named, dated studies rather than repeated claims.

What AI search visibility actually is

AI search visibility is whether AI answer engines such as ChatGPT, Google AI Overviews and Perplexity read, trust and cite a page when generating an answer.

Generative engine optimisation, usually shortened to GEO, is the discipline of making a page legible and trustworthy enough that an AI answer engine cites it directly inside a generated answer, rather than sending a visitor through a list of ten blue links. We define the term in full in what GEO actually means, and this guide is the operational layer on top of that definition: the levers that move citation, in the order that actually earns them.

GEO does not replace search engine optimisation. It sits alongside it, because the crawlers behind ChatGPT, Gemini and Perplexity still depend on the same crawlable, well structured HTML that search engines have rewarded for two decades. Where the two disciplines diverge is in the unit being optimised: SEO earns a ranking position, GEO earns a sentence inside someone else's answer. We set out where the two overlap and where they do not in SEO is not dead, it is being joined by GEO.

The distinction matters operationally. A page can rank first on Google and still be invisible to an AI engine if the delivery layer breaks, which is the first lever this guide covers. It also matters for budget: teams that treat GEO as a content problem keep publishing, see no movement, and conclude the channel does not work for them. Teams that treat it as a delivery and evidence problem find the actual blockage, usually in an afternoon.

For a B2B site, the stakes are higher than they look. A buyer researching a category now often asks an AI engine to compare vendors before they ever visit a site directly. If that engine cannot read or trust a page, the page is absent from a comparison it would otherwise win on merit. Visibility here is not a vanity metric, it decides which vendors get considered at all.

Why this is not a fringe channel any more

OpenAI reported 900 million weekly ChatGPT users on 27 February 2026, more than double the 400 million reported a year earlier, putting AI answers in front of buyers by default.

On 27 February 2026, OpenAI reported more than 900 million weekly active users for ChatGPT, more than double the 400 million it reported in February 2025. That growth curve is the whole argument for treating AI search visibility as infrastructure rather than an experiment: a meaningful share of the research a buyer does before they ever fill in a form now happens inside a chat window, not a search results page.

We track what that shift is worth in referral terms, and where it already beats traditional organic search for us, in AI traffic already outperforms Google for us. The pattern holds across the categories we operate in: fewer visits, a higher proportion of them from someone who already has intent, because the AI engine did the filtering before the click happened.

None of this makes traditional search obsolete. It makes the AI answer layer a second gate a page has to pass, with its own delivery and citation mechanics, on top of the ranking gate that already existed. A page that clears the first gate and fails the second is invisible to a growing share of buyers regardless of where it sits in a traditional results page.

The consequence for a marketing or revenue team is a new line item, not a new department. The skills involved, crawlability, structured content and defensible data, sit close to existing SEO and content operations. What changes is the checklist, and getting the checklist wrong by copying 2024 or 2025 SEO advice wholesale is the single most common mistake we see, because the mechanics genuinely differ.

The six levers that actually earn citations

Our own August 2026 audit found six levers that move AI citation: crawlability, correct delivery, structure, schema, freshness and original data.

We ran a full AI search visibility audit of our own site in August 2026, deliberately as a first party exercise rather than a vendor report, because we wanted findings we could stand behind rather than repeat. Six levers came out of it, and volume was not one of them:

The first two levers are binary gates. A page either can be reached and correctly parsed, or it cannot, and no amount of quality further down the list compensates for failing either one. The middle two, structure and schema, are legibility levers: they decide how cheaply an engine can extract an answer once it has the page. The last two, freshness and original data, are the levers that decide whether the page is worth extracting from in the first place, and they are the two hardest for a competitor to copy without doing the same underlying work.

The rest of this guide takes each lever in turn, starting with the one most sites get wrong first.

Crawlability: the lever most sites get wrong first

A page an AI crawler cannot reach or correctly parse cannot be cited, whatever the content says, and the most common failure is a robots.txt or delivery mistake.

The three big AI labs no longer run a single crawler each. OpenAI now runs GPTBot for model training, OAI-SearchBot for its search index, and ChatGPT-User for the live fetch that happens when someone clicks a citation link. Anthropic runs the equivalent split with ClaudeBot, Claude-SearchBot and Claude-User. A robots.txt file now has to make at least three separate decisions per AI company: training, indexing and live retrieval, not one.

The mistake we see most often, and found on our own site during the August 2026 audit, sits in how robots.txt groups work. Under RFC 9309, the standard that governs robots.txt, a named user agent group does not inherit from the wildcard group. A crawler matches exactly one group. A robots.txt with a wildcard group disallowing /admin/ and /api/, and a named GPTBot group that only allows crawling, gives GPTBot free run of /admin/ and /api/, because the named group replaced the wildcard rather than adding to it. We had that exact defect and fixed it the day we found it, by repeating every disallow rule inside every named crawler group rather than relying on the wildcard to cover them.

Correct delivery is the second half of crawlability, and it fails for a reason almost nobody checks: the Content-Type header. We found a page on our own site that rendered perfectly in a browser but was served with a header that told an AI crawler it was not HTML at all, making it functionally invisible to citation even though a human visitor would never notice a problem. The full mechanism and the fix are in the header that makes perfect HTML invisible to AI search.

Neither defect shows up in a normal browser check, which is exactly why both survive unnoticed for months. Both only show up when you request the page as the crawler requests it, with the crawler's own user agent string, and read the response the way a machine reads it rather than the way a browser renders it.

Structure and schema matter less than the hype claims

The evidence on schema markup is mixed at best. Ahrefs found a small decline in AI Overview citations after adding schema, and a 2026 analysis found organic rank was the stronger predictor.

Most 2026 GEO advice treats schema markup as close to mandatory, and some of the figures used to justify that claim do not trace to a named source when checked, which is exactly the trap our own fact gate exists to catch. Two studies do trace cleanly to a named organisation and a named methodology, and neither supports the strong version of the claim.

Ahrefs tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026, matched against 4,000 control pages that did not add it. Citations in Google AI Overviews fell 4.6% on the treated pages relative to controls, a small but statistically real decline. The effect on Google AI Mode and ChatGPT citations was too small to distinguish from noise in either direction, on pages that already had over 100 AI Overview citations each before the test began.

A separate 2026 working paper by researcher Kurt Fischman, published on SSRN, collected 730 AI citations from ChatGPT and Gemini across 75 commercial queries spanning five categories. Schema presence looked negatively associated with citation in the first pass, but that effect collapsed once the analysis controlled for Google's organic ranking position, which independently enriches schema prevalence among top ranked pages regardless of whether an AI engine ever cites them. The predictor that survived the correction was organic rank itself: pages in position one were cited in 43% of the queries where they appeared, falling to 5% by position seven.

None of this makes schema worthless. FAQ and Article schema still give an engine a clean, structured answer to lift, which is why we still use it on every page, and it still helps a conventional search engine understand a page and qualify it for a rich result. It means schema is not a citation lever on its own, and budget spent chasing schema as the fix is usually budget better spent on the ranking and structure work covered in how to actually get cited by AI search engines. Structure without schema still gets cited. Schema without structure and rank rarely does.

Freshness and original data are the levers competitors cannot copy cheaply

The lowest effort GEO fix, llms.txt, shows no measurable citation benefit at scale. Freshness and genuinely original data are harder to fake and harder for a competitor to copy.

llms.txt is the file most GEO checklists recommend first, because it is the cheapest thing to add. SE Ranking's May 2026 analysis of roughly 300,000 domains found 10.13% adoption, and no measurable relationship between having the file and how often a domain gets cited, once site authority, schema density and content recency were controlled for. Among the top 1,000 domains by traffic, adoption was effectively zero. We keep the file on our own site because it costs nothing and is not wrong to have. We do not treat it as a lever, and we would not recommend a client spend a sprint on it.

Freshness and original data earn citation for a duller reason: they are expensive to fake. A dated, verifiably updated page signals to an AI engine's retrieval layer that the answer is current, and a page carrying a number, a finding or a position nobody else has published gives the engine something to cite that a summary of the top ten results cannot replicate. That is also why brand and entity naming compound over time rather than resetting with every algorithm update, a pattern we cover directly in the name an AI recommends is not an accident.

Freshness has a failure mode worth naming directly: the silent edit. A page gets updated, the visible copy changes, but the dateModified field in its schema does not move, so the page reads as stale to anything checking the structured date rather than the rendered text. We check both fields separately in our own audit process for exactly this reason, because they drift apart more often than either changes alone.

The practical implication is that a GEO programme built entirely around technical fixes plateaus. The technical work removes the ceiling. Original data and a defensible point of view are what a page actually gets cited for once the ceiling is gone, which is also why this engine exists: a dated piece with a first party finding, published weekly, is a freshness and originality signal a competitor cannot replicate by copying the checklist.

How to run your own AI search visibility audit

Run the six levers in order: confirm crawler access, check delivery headers, test structure, validate schema, log freshness, and inventory what on the page is genuinely original.

The order matters. Fixing schema on a page an AI crawler cannot reach changes nothing, and validating structured data on a page served with the wrong Content-Type header is wasted work until the delivery fault is fixed. Run the six levers in the sequence below, on the pages that matter most first, and use our full GEO audit checklist for the detailed pass on each one once this sequence has surfaced where the blockage actually is.

Most sites find their first real defect in step one or step two. Very few sites need to get past step four before finding something worth fixing, which is itself a useful signal: if an audit reaches original data inventory without finding a crawlability or delivery problem, that site is already ahead of most of the market on the levers that gate citation entirely.

None of the six steps needs specialist tooling. A terminal, a way to set a custom User-Agent header on a request, and a JSON validator cover the first four. The last two are editorial judgement rather than technical checks: does the update date on this page reflect a real change, and could a competitor publish this exact page tomorrow. If the honest answer to the second question is yes, the page is structurally sound and still not finished.

Frequently asked questions

What is AI search visibility?

AI search visibility is whether an AI answer engine such as ChatGPT, Google AI Overviews, Perplexity or Claude reads, trusts and cites a page when it generates an answer, rather than the page simply ranking in a traditional results list.

Does llms.txt improve AI citations?

The evidence says no. SE Ranking's May 2026 analysis of roughly 300,000 domains found no measurable citation benefit from having an llms.txt file, once site authority, schema density and content recency were controlled for, and only 10.13% of sites had added one.

Does adding schema markup increase AI citations?

Not reliably on its own. Ahrefs tracked 1,885 pages that added schema between August 2025 and March 2026 and found a small decline in AI Overview citations relative to controls, and a separate 2026 SSRN analysis found organic rank position, not schema, was the strongest predictor of citation.

Why do AI crawlers sometimes see different content to what a browser shows?

Usually a delivery fault, such as a robots.txt named group that overrides the wildcard group instead of inheriting from it, or a Content-Type header that misdescribes the page. Both can make a page technically unreachable or unreadable to a crawler even though it renders fine in a browser.

How many AI crawlers does a robots.txt file need to account for in 2026?

At least three per major AI company: a training crawler such as GPTBot or ClaudeBot, a search indexing crawler such as OAI-SearchBot or Claude-SearchBot, and a live user fetch crawler such as ChatGPT-User or Claude-User, each needing its own explicit rules under RFC 9309.

What matters more than schema for getting cited by AI?

Crawlability and correct delivery come first, because a page an engine cannot reach or correctly parse is never cited regardless of its content. After that, freshness and genuinely original data, a fact or position that does not appear on a competitor's page, matter more than schema markup on its own.

How often should an AI search visibility audit be repeated?

Repeat the crawlability and delivery checks whenever the site, CDN or hosting configuration changes, since those are the checks most likely to regress silently. Review freshness and original data on a slower cycle, because those levers move with the content calendar rather than with infrastructure changes.