AI search tactics don’t port cleanly across models, analysis finds

Why universal AI search playbooks keep breaking

The biggest mistake in AI search right now is treating it like classic SEO with a fresh coat of paint. OpenAI, Google, Microsoft, and Anthropic are all exposing different mixes of web search, citations, retrieval, and answer generation, so a tactic that helps one system can do almost nothing in another. That is the practical warning here: if you borrow a trick from one model and assume it ports cleanly, you are probably optimizing for the wrong surface.

llms.txt is where the illusion starts to crack

The /llms.txt proposal makes the problem easy to see. It says the file should provide brief background information and guidance, plus links to more detailed markdown files, to help LLMs use a website at inference time, and its own background says the motivation is simple: context windows are too small to handle most websites in full. That sounds tidy on paper, but it only works if the target system actually reads, requests, and weights that file the way you expect.

That is also why the old SEO habit of looking for one clean technical switch is so misleading in AI search. Google’s documentation says robots.txt is mainly for crawler control, not for keeping a page out of Google, and it says noindex is the correct way to keep a page out of search results. In traditional SEO, those levers are clear enough to feel universal; in LLM land, they are just pieces of a much messier visibility puzzle.

John Mueller’s comparison of llms.txt to the old meta keywords tag landed because it captured that skepticism in one line. The important takeaway is not that every machine-readable file is useless, but that one ecosystem’s shrug tells you very little about what another model will ingest, ignore, or surface later. If you treat that skepticism as a universal verdict, you end up missing the platform-specific behavior that actually matters.

The platforms are not interchangeable

OpenAI says ChatGPT search can return timely answers with links to relevant web sources, and it will search the web automatically when a query might benefit from web information. Google says AI Overviews rolled out to everyone in the U.S. in May 2024, then expanded to more than 100 countries and territories and to more than 1 billion monthly users, while Search Central frames AI features from a site-owner perspective and says SEO best practices remain relevant. Microsoft positions Copilot Search as quick summarized answers with cited sources and suggestions for further exploration. Those are three different products with three different citation experiences, not one AI search system with three logos on top.

Google’s own guidance makes the point even sharper. It says AI features like AI Overviews and AI Mode have technical requirements, visibility controls, and performance measurement considerations, and it also says generative AI content that is scaled without added value may violate spam policy. So the game is not “produce more AI content and hope it ranks everywhere”; it is “make something useful enough that each system has a reason to cite or surface it.”

What to test instead of relying on a universal playbook

The right move is to stop thinking in slogans and start thinking in separate experiments. A change that improves citation frequency in ChatGPT search may not budge Google AI Overviews, and a page structure that helps Copilot Search find a cited answer may not matter to a model that leans harder on its own retrieval stack or safety layer. Even Claude’s Research workflow, which can search across the web, Google Workspace, and other integrations, shows how different the underlying systems can be once you move beyond a single search box.

A sane validation process should look like this:

• Track each platform separately. Measure ChatGPT search, Google AI Overviews or AI Mode, and Copilot Search as distinct surfaces, because each one presents citations and outbound links differently.

• Test one change at a time. If you alter headings, schema, internal links, or page depth, watch whether the effect shows up in one ecosystem, none of them, or all of them. Google’s docs make clear that AI features still ride on core SEO fundamentals, so the cleanest tests are the ones that isolate a single variable.

• Use the right control for the right job. robots.txt is for crawl access, noindex is for exclusion, and mass-produced generative content without value is a bad bet no matter how fashionable the tactic sounds.

• Keep a platform log. Record which queries trigger citations, which pages get referenced, and where your content appears without a click. That is the only way to catch the pattern shifts that get lost when teams talk about “AI search” as if it were one channel.

The bottom line

AI visibility is not a single surface, and it is not governed by a single playbook. The teams that will do well here are the ones that treat ChatGPT, Google, Copilot, and Claude as separate systems, validate each one on its own terms, and stop pretending that old SEO assumptions travel intact into an ecosystem still changing under their feet.