Search traffic was never the only distribution channel, but it used to be the predictable one. Answer engines changed the terms: a model summarizes your positioning, cites one competitor, and never sends the visit. The sites that get cited are not the loudest—they are the ones whose facts were legible when the crawl happened.
AI crawler access is a policy decision, not a side effect
robots.txt now carries a business decision. GPTBot, ClaudeBot, PerplexityBot, and OAI-SearchBot can be allowed, blocked, or allowed selectively—and the default many teams ship with is accidental. Decide per environment too: preview and staging deployments that disallow everything are correct; production that inherits the same rule is invisible. The audit starts with reading what you actually serve, not what you assume.
llms.txt is a handshake, not a ranking factor
A plain-text llms.txt file listing what the site is, what the main pages are, and what each one contains costs almost nothing to maintain. It is not a standard yet and it will not move rankings. It earns its place because retrieval pipelines and research agents read it when they need a fast, honest map of a domain—one paragraph of description per page beats scraping twenty templates to reconstruct the same information.
Structured data decides attribution
When an answer engine quotes a company, it needs something to attach the quote to. Organization, Service, FAQPage, and BlogPosting markup with dates, authors, and canonical URLs give models stable handles on entities. Sites without them get paraphrased; sites with them get named. The difference shows up in whether the generated answer carries your brand or a category description.
Answer-shaped content gets quoted
One question, one direct answer, then the detail. Sections that open with a claim a model can lift verbatim—before the caveats and the backstory—are the ones that survive summarization. That is not writing for machines at the expense of readers; it is the same discipline good technical writing always demanded, now enforced by a harsher editor.
Measure visibility where answers are generated
Rank trackers do not cover this surface. Periodic prompts against the engines your buyers actually use—"who does X for Y in region Z"—plus a log of which pages your own support questions link to, give a workable signal. Treat the results like an incident dashboard: when the answer stops naming you, something on the exposure checklist changed.