Invisibility is decided before the model writes a word
When ChatGPT answers a current-state question about your category, the model is not recalling your company. It is reading a handful of documents fetched seconds earlier and writing a paragraph from them.
That distinction is the whole game. If your page is not in the retrieved set, no amount of persuasive copy on it matters. You are not losing an argument with the model. You are not in the room.
OpenAI documents the mechanism only loosely, and the vagueness is worth knowing. The API documentation states the model “can choose to search the web or not based on the content of the input prompt”. The help centre says search “may” activate when a question would benefit from current information. Neither publishes a threshold, a classifier or a rule. Anyone who tells you exactly when ChatGPT decides to search is describing a guess.
What OpenAI does document is more useful than it looks. The user's question is rewritten into several targeted queries before being sent to search partners, and the system distinguishes citations — the subset shown inline — from sources, the full list of URLs consulted. So there are two separate ways to lose: never being retrieved, and being retrieved but never quoted.
Insight
The second failure is the larger one. An AirOps analysis of 548,534 retrieved pages across 15,000 prompts found roughly 85% were never cited. Ahrefs, analysing 1.4 million prompts, put the citation rate near 50%. Different samples, different methods, never reconciled. Take the direction, not the number.
What ChatGPT actually cites
Every large-scale study of AI citation behaviour published in 2025 and 2026 comes from a company that sells AI-visibility tooling. That does not make the data wrong. It does mean you should read it as evidence rather than scripture — and the gaps between the studies are as informative as the headlines.
Two things survive across all of them. Concentration: a small number of domains account for a large share of citations in any category. And volatility: that set is not stable. Profound measured citations per answer falling by a tenth after a single model update, and unique cited domains falling in 56 of 58 industries at the same time.
A visibility position built on one source is a position you can lose in a week without anyone telling you.
The August 2026 Reddit collapse, and what it should have taught everyone
In mid-August 2026 Reddit's share of ChatGPT Search citations fell off a cliff. Promptwatch measured it at 3.83% dropping to 0.52% in roughly four days — an 86% decline. Otterly, measuring per-brand across sixteen tracked reports over an overlapping window, put the fall at “at least 73%”, and noted official and reference sites gaining share as Reddit lost it.
Three details matter more than the headline. The numbers disagree: 86% and 73% describe the same event measured two ways, and averaging them would produce something meaningless. Promptwatch flagged its own finding as provisional and said it could not rule out a data-collection fault — the vendor with the most dramatic number was the most careful about it. And nobody knows why it happened: OpenAI did not comment, the site-operator change people pointed at was six days earlier, and a similar collapse occurred in September 2025 for reportedly unrelated reasons.
The lesson is not that Reddit is finished. It is that a citation source is infrastructure you do not control, and a visibility strategy resting on one of them is one model update away from zero.
The advice with no evidence behind it
This is the section most guides in this category leave out, so it is worth being blunt.
llms.txt. Rankability, a company that sells SEO software, reported on 18 September 2026 that 830 of the Tranco top 10,000 domains, 8.3%, confirmed-serve an llms.txt or llms-full.txt file; another 4,211 returned an inconclusive result, so the true rate is uncertain. Other counts differ: 5.61% of the top 10,000 origins in HTTP Archive data in June 2026 (an analysis by Casey Burridge, 20 June 2026), and 6.6% of 1,000 top origins in a single crawl on 1 to 2 October 2026 (LintLab, which sells a sitemap checker). Different lists, different dates, same message: it is rare. On use, Ahrefs, which also sells AI-visibility tools, examined 137,210 domains for May 2026 and reported on 15 June 2026 that 97% of llms.txt files received no requests at all; it notes its sample skews technical and that a fetch does not prove a bot read or used the file. What has never been published — by OpenAI, Google, Anthropic or Perplexity — is any confirmation that these files are read at all. The adoption data measures how many people believed it would work, not whether it does.
Schema markup. Google's John Mueller, asked directly whether schema helps LLMs understand entities, answered “yes, no, and it depends”, and characterised some markup as wishful thinking. He was explicit it was a personal opinion, not guidance. The one quasi-controlled test we could find is Otterly's own, on its own site: five schema types across four URLs, tracked over three months and 319 prompts. The results ran in every direction at once — ChatGPT down 71%, Gemini down 35%, Perplexity unchanged, AI Overviews up 611% — and the authors say the lift coincided with a broader algorithmic shift and that causality is unclear.
“Brand mentions beat backlinks.” Ahrefs' study across 75,000 brands found YouTube mentions correlated more strongly with AI visibility than backlink volume. That is a correlation in observational data. It does not show that acquiring mentions causes visibility to rise, and the study does not claim it does.
Insight
None of this means schema, structured data or brand mentions are worthless. It means the confident numbers you have been shown about them do not exist yet. Anyone selling you a tactic here on the strength of settled evidence is selling you something else.
What the evidence actually supports
Strip out everything unverified and a short list remains.
Content characteristics affect citation, and this one is peer-reviewed. The KDD 2024 Generative Engine Optimization paper found adding quotations improved visibility by roughly 41%, statistics by 30–40% and cited sources by around 30%, while keyword stuffing did almost nothing — validated against real Perplexity queries at 22–37% on a smaller secondary sample. The authors note the primary testbed was a custom engine, not a production pipeline.
Structure helps, but not the way the checklists say. An April 2026 analysis of 21,143 citations across 602 prompts found deeply-cited pages carried 11.4 times more words and 12.5 times more headings than rarely-cited ones — and that Q&A formatting on its own had a slightly negative effect of −5.7%. The authors argue structure only helps when it carries genuine evidence: definitions, statistics, comparisons, procedures. The most repeated advice in this space, “add an FAQ section”, is now contradicted by data.
Being fetchable is not optional. ChatGPT's crawlers, Perplexity's PerplexityBot and Anthropic's Claude-SearchBot all have documented user agents and robots.txt behaviour. A surprising number of companies block the engines they want to be recommended by. It is the cheapest thing on this list to check and the most embarrassing to get wrong.
Comparative material that names competitors. Engines answering “which is better for X” need documents that make the comparison explicitly. If only your competitors publish those, only they get quoted.
How to know whether any of it worked
Engines are non-deterministic. The same question asked twice, minutes apart, can return a different set of sources. One flattering answer is not evidence, and a screenshot is not a measurement.
What holds up is dull: a fixed set of buying-intent questions, fielded the same way on each pass, scored on whether you were mentioned and which domains were cited rather than on the wording of the reply, and compared against a baseline taken before you changed anything.
If you have not measured the before, you cannot claim the after.
OpenAI web search tool documentation and ChatGPT search help centre (undated; read 4 Oct 2026) · Semrush, The Most-Cited Domains in AI (10 Nov 2025) · Ahrefs, Why ChatGPT Cites One Page Over Another (15 Apr 2026) · AirOps (12 Mar 2026) and ALM Corp (14 Mar 2026) retrieval analysis · Profound Index Report, Summer 2026 (19 Aug 2026) · Promptwatch finding via Search Engine Land (19 Aug 2026) and Axios (20 Aug 2026) · Otterly, ChatGPT cut Reddit citations by at least 73% (27 Aug 2026) · Search Engine Journal (19 Aug 2026) · Rankability, llms.txt Adoption Research Report (18 Sep 2026) · Burridge, State of llms.txt adoption (20 Jun 2026) · LintLab via PPC Land (2 Oct 2026) · Ahrefs, llms.txt study (15 Jun 2026) · Search Engine Roundtable (2 Jan 2026) · Otterly, GEO Experiment: Does Schema Markup Really Impact AI Search? (23 Mar 2026) · Aggarwal et al., GEO, KDD 2024 (arXiv:2311.09735) · Zhang, He & Yao (arXiv:2604.25707, Apr 2026) · Ahrefs, 75,000-brand study (12 Dec 2025, updated 12 Aug 2026)