An AIO Roadmap

Why a roadmap rather than a checklist

AI assistants and AI search surfaces answer many questions by retrieving material at the time the question is asked and composing an answer over what they retrieved. The gates on that retrieval are documented in places, and they are not one gate. Google states that to be eligible for its generative AI features a page must be indexed and eligible to be shown in Google Search with a snippet. OpenAI documents a separate agent, OAI-SearchBot, and states that allowing it in robots.txt is what helps a site appear in ChatGPT search results. Two systems, two independent conditions, and neither of them is an explanation of how anything gets chosen.

A roadmap is more useful than a checklist here because the work is dependent rather than parallel. A page that cannot be reached cannot be read. A page that can be read but describes its subject inconsistently is harder to represent accurately. An accurate page still leaves an entity thinly supported if nothing off that domain corroborates it. And none of it can be managed without some way to observe what is happening.

AIOFacts proposes five stages: reachability, legibility, evidence, corroboration, measurement. The evidence label on this piece is position, and it is meant literally. Each mechanism below is sourced. The ordering of them is our judgment about dependency, not a documented behavior of any AI system.

What the mechanism is, and what nobody outside can see

Three things are publicly observable. First, access: operators publish named user agents and, in OpenAI's case, IP ranges at endpoints such as searchbot.json and gptbot.json, so a site owner can verify from their own logs which agent arrived. Second, the stated conditions of eligibility, such as Google's indexed and snippet-eligible requirement. Third, the output itself, which anyone can prompt and record.

What is not observable is the middle. How a given system selects among retrieved candidates, what it weighs, and how it decides to attribute are internal to the operator. AIOFacts does not describe those internals and treats any confident public account of them as unsupported. Platform behavior also varies between assistants, and a mechanism documented by one operator says nothing about another.

The gap between presence and accuracy is documented. The Tow Center for Digital Journalism tested eight generative search tools across 1,600 queries in March 2025 and reported that collectively they answered incorrectly on more than 60 percent of queries, with results ranging from a 37 percent error rate for the free version of Perplexity to 94 percent for Grok 3. Appearing in an answer and being described correctly in it are separate outcomes, and the second is the one this roadmap is aimed at.

Stage one: reachability, and the training versus answering distinction

Reachability is first because it is the only stage where a single configuration error removes every later possibility. It is also the stage most often misconfigured in a way that looks deliberate and was not.

OpenAI documents four agents with different jobs. GPTBot crawls content that may be used in training foundation models. OAI-SearchBot surfaces sites in ChatGPT's search features. ChatGPT-User handles user-initiated actions, and the documentation notes that robots.txt rules may not apply to it because a person triggered the fetch. OAI-AdsBot only visits pages submitted as ads. Anthropic documents ClaudeBot as its general-purpose crawler, states that it follows robots.txt, and specifically advises against blocking by IP address, because doing so impedes the crawler's ability to read the robots.txt file that expresses the preference in the first place.

The practical consequence is that a blanket disallow aimed at training can also remove a site from answering surfaces it wanted to be in. Cloudflare's Content Signals Policy addresses the same problem by naming three separate uses in robots.txt: search, meaning building a search index, ai-input, meaning feeding content into a model for a real-time answer, and ai-train, meaning training or fine-tuning. Cloudflare states plainly that content signals express preferences and are not technical countermeasures against scraping.

The same distinction is being standardized. The IETF AI Preferences working group's draft, A Vocabulary For Expressing AI Usage Preferences, was at revision 06 dated 28 April 2026, with an intended status of Proposed Standard. It defines two usage categories, AI model training and search, each taking a value of allowed, disallowed, or unknown. It is an active Internet-Draft, not a finished standard, and should be described that way.

• Know which named agents you allow and which you disallow, per agent, deliberately.
• Treat training access and answering access as two decisions, not one.
• Verify from server logs and published IP ranges rather than trusting a user agent string.
• Exit test: you can state your position for each named agent, and your logs agree with it.

Stage two: legibility, which is mostly ordinary work

Legibility is where the most confident advice in the market is least supported, so it is worth quoting the platform guidance directly. Google states that no specific optimization is required for AI Overviews and AI Mode, that you do not need to create new machine readable files, AI text files, markup, or Markdown, that there is no requirement to break content into small pieces, and that structured data is not required for generative AI search. Its stated position is that generative AI features are rooted in core Search ranking and quality systems, so ordinary search best practices continue to apply.

That does not make structured data pointless. It makes it a documented mechanism for specific search features rather than a documented lever on AI answers, and the distinction matters when deciding where effort goes. AIOFacts treats structured data as one of its seven working signals, on the reasoning that machine-readable descriptions reduce ambiguity, not on any claim about what a model does with them.

The work that belongs in this stage is unglamorous: one clear canonical description of the entity, the same name and the same core claims wherever they appear, pages that state what they are about early and plainly, and no dependence on client-side rendering for content that matters. This may help reduce the number of conflicting descriptions a system has to reconcile. It is not a guarantee that any system reconciles them in your favour.

Stage three: evidence attached to claims

The one controlled experiment commonly cited on this point is GEO: Generative Engine Optimization, presented at KDD '24 by Aggarwal and colleagues, which introduced a benchmark of 10,000 queries and reported that content-level changes such as adding citations, quotations, and statistics improved visibility in generative engine responses by up to 40 percent, while classical keyword-density signals showed little effect.

The limits of that result deserve equal weight. It is one study, run against a specific set of generative engines at a specific time, on a constructed benchmark rather than a live commercial corpus, and the systems it tested have changed since. AIOFacts treats it as supporting evidence for a direction, not as a rule about how AI systems behave, and would treat any campaign built on the number 40 percent as overreading it.

There is a second reason to attach evidence to claims that does not depend on the study at all. A sourced claim is checkable by a human reader, survives being quoted out of context, and gives anyone verifying an answer somewhere to land. That value is independent of whether it also affects retrieval.

Stage four: corroboration you do not control

The Tow Center study included a finding that matters for planning. ChatGPT cited syndicated versions of USA Today articles despite USA Today blocking ChatGPT's crawler, and licensing arrangements did not guarantee accurate attribution: the San Francisco Chronicle, whose parent had a partnership with OpenAI, was correctly identified in one of ten queries.

Read carefully, that is a statement about copies. Material about an entity circulates through syndication, aggregation, directories, and third-party writeups, and those copies are not governed by the entity's own robots.txt. Controlling your own domain is necessary and is not sufficient.

The stage-four work is therefore about consistency across sources you did not publish: accurate profiles on the platforms that describe you, corrections requested where a third-party description is wrong, and the same core facts in each place. AIOFacts calls this citation authority and brand recognition in its working framework. Neither is a claim that a model counts mentions. Both are a bet that a description repeated accurately in independent places is harder to get wrong than one that exists in a single location.

Stage five: measurement, which is the weakest link

Google announced Search Generative AI performance reports in Search Console in June 2026, which gives one first-party view of performance in generative AI features for one platform. Server logs give a second view: which named agents fetched which URLs, verifiable against the published IP range files rather than against a user agent header that anyone can set.

Beyond that, the available methods are observational. A prompt can be run repeatedly, across assistants, and the answers recorded, which produces a distribution rather than a fact. Outputs vary between runs, between accounts, and over time, so a single answer is one observation and should be logged as one. Where an answer is wrong about an entity, the wrongness is the finding worth keeping, because it is specific and actionable in a way that a visibility score is not.

There is no cross-platform equivalent of Search Console for most assistants, and no public method for establishing why any particular answer was produced. Any measurement practice that claims otherwise is claiming access it does not have.

Sequencing, exit tests, and what this roadmap does not promise

The stages are ordered by dependency, and each one has a test that can be failed. Reachability: your agent-by-agent position is deliberate and confirmed in logs. Legibility: an unfamiliar reader can state what the entity is from a single page, and the same description holds across the site. Evidence: every non-obvious claim carries a source a stranger can check. Corroboration: the third-party descriptions you can find are accurate, and the wrong ones have been challenged. Measurement: you are recording observations over time rather than reacting to a single screenshot.

What none of this produces is a guarantee. No stage guarantees appearance in any AI answer, no stage guarantees accurate attribution, and the terminology for the whole activity is unsettled: AIO, GEO, AEO, and LLMO are all in use, and AIOFacts uses AIO as its own working term while treating the others as narrower framings rather than settled subsets. Anyone promising a ranking in an assistant is promising something nobody outside the operator is in a position to deliver.

The honest version of the objective is narrower and still worth the work: make it easy for a system to reach the material, hard to misread it, and possible for anyone, human or machine, to verify it against something other than your own assertion.