Source cutoff: July 16, 2026

The useful answer is less exotic than the sales pitch.

Google's current guidance says optimization for its generative search experiences is still SEO. The systems use Google's Search index, retrieval, and related-query techniques to assemble answers. A page still has to be crawlable, indexable, useful, and eligible to appear in Search. Google does not require a special AI file, special AI schema, microscopic content chunks, or pages for every wording variation.

OpenAI, Perplexity, and Anthropic each document crawler controls with different jobs. Search crawlers, training crawlers, and user-triggered agents are not the same policy decision.

So I would prepare a local-service website in four layers:

  1. Eligible — the intended search system can access and index the page.
  2. Explicit — the service, area, fit, proof, price or estimate path, and next step exist in visible text.
  3. Citable — the page has a direct answer, named source or accountable author, evidence, scope, date, and honest limitations.
  4. Useful after the click — a person can understand the offer and complete the next step.

That is a website standard, not a promise that an assistant will cite or recommend the business.

AI search is not a separate content universe

Google describes query fan-out: its AI search features may issue multiple related searches and combine supporting pages into an answer. That changes how broad the research journey can become, but it does not make one page for every possible subquery a good strategy.

Google's July 2026 guidance warns against scaled pages made for query variations and prioritizes unique, expert-led, non-commodity material. The practical implication is not “publish more.” It is “own a useful answer.”

For a service-business website, that means:

  • commercial pages own stable facts about the offer, fit, price, process, area, and next step;
  • short FAQs remove friction on the page that already owns the subject;
  • complete articles explain consequential decisions, mechanisms, costs, failures, and tradeoffs; and
  • reference assets make a method, standard, checklist, calculation, scorecard, or benchmark reusable.

A question does not deserve a new URL merely because a buyer or an AI system could ask it.

Layer 1: Eligible

Start with ordinary retrieval.

For Google, the page must be eligible for Search and able to display a snippet. Important content should be available as text. Canonicals, internal links, sitemaps, rendering, and accidental noindex or access blocks still matter.

For ChatGPT search, OpenAI documents OAI-SearchBot. It documents GPTBot separately for potential model training and ChatGPT-User for user-triggered visits.

Perplexity similarly documents PerplexityBot for search and Perplexity-User for user requests. Anthropic documents Claude-SearchBot, Claude-User, and ClaudeBot separately.

That creates three policy rows, not one “allow AI” switch:

Purpose Examples Decision
Search retrieval Googlebot, bingbot, OAI-SearchBot, PerplexityBot, Claude-SearchBot Allow or block based on whether the site wants to be eligible for those search experiences.
User-triggered retrieval ChatGPT-User, Perplexity-User, Claude-User Decide whether people may use those products to access the public site on their behalf.
Model training GPTBot, ClaudeBot, Google-Extended and other documented controls Make a separate intellectual-property and exposure decision. Search access does not silently settle it.

Crawler access only establishes a technical possibility. It does not prove crawling, indexing, citation, recommendation, or training inclusion.

Layer 2: Explicit

Ask whether the important facts can be pointed to in visible text:

  • What service is provided?
  • Who is it for?
  • Where or how is it delivered?
  • What conditions change fit, scope, or price?
  • What proof can the reader verify?
  • What happens after the next step?
  • Who is responsible for each part?

Do not make a logo, slogan, image, video, PDF, structured-data property, or business profile carry a fact the page itself never states.

Use applicable structured data to reinforce visible facts. Google's policies say structured data can make a page eligible for supported features, but it does not guarantee display. Unsupported or invisible claims do not become credible because they are wrapped in JSON-LD.

Layer 3: Citable

A page becomes easier to cite accurately when a reader can lift the answer without changing its meaning.

I would look for:

  • a direct answer near the beginning;
  • explicit scope and conditions;
  • a named author or accountable owner;
  • primary sources for volatile platform facts;
  • original method, calculation, demonstration, checklist, or evidence;
  • reviewed date and source cutoff;
  • limitations and counterexamples; and
  • a stable URL with a clear relationship to the rest of the site.

This is not content “chunking.” It is ordinary editorial clarity.

The part that becomes hard to copy is the useful part: an original acceptance standard, a transparent proposal scorecard, a repeatable lead-capture test, a reproducible benchmark, or permissioned evidence showing what was actually inspected and verified.

Layer 4: Useful after the click

Search visibility is unfinished when the person who arrives still cannot decide or act.

Make the next step explicit and test it:

  • call during stated hours;
  • request an estimate;
  • book a specific service;
  • send project details;
  • compare scope and price; or
  • inspect the evidence behind a claim.

Use semantic links and buttons, visible labels, accessible names and states, stable layouts, understandable errors, and accurate success feedback. A site that confuses a person is not “agent ready” merely because it has schema.

What not to buy as an AI-search requirement

Claim What current primary guidance supports
“You need special AI schema.” Google says no special markup is required. Use supported structured data that matches visible content.
“You need an llms.txt file.” Google says no special AI file is required. Maintain one only for a documented consumer; it is not a ranking promise.
“Turn every long-tail question into a page.” Google warns against scaled query-variation pages. Use one owner page per distinct reader job.
“FAQ schema makes assistants cite you.” Useful FAQs can help readers, but markup is not a general AI-citation lever.
“Let every AI crawler in.” Search, user-triggered access, and training are separate policies.
“Structured data makes the page authoritative.” Markup describes visible facts. It does not create experience, evidence, or independent corroboration.
“Track your universal AI rank.” Assistant answers vary by product, time, location, personalization, and wording. Use a fixed prompt sample as a directional observatory.
“Mentions anywhere build model memory.” Retrieval-time citations and future training are different systems. Artificial mentions are neither trustworthy nor controllable.

How to measure the work

Use the native evidence the platforms expose.

Google began rolling out a generative-search performance report in Search Console in 2026. Bing Webmaster Tools now has AI Performance reporting for citations, cited pages, and sampled grounding queries. Availability can vary by property.

Also track:

  • index coverage and crawl access for priority pages;
  • AI-search referral landing pages and qualified actions;
  • independent mentions and links from relevant sources;
  • the percentage of priority pages meeting the citation-grade standard;
  • stale volatile claims and update time; and
  • a fixed monthly prompt panel labeled as a sample, not a rank tracker.

Set improvement targets only after the first measured baseline exists.

A fifteen-minute local-service website check

  1. Find the page that owns the main service and area. Can a stranger state both?
  2. Read the page without relying on images, video, profiles, or schema. Are the important facts still present?
  3. Check status, canonical, internal path, sitemap, rendered text, and crawler policy.
  4. Compare visible facts with applicable structured data and current business profiles.
  5. Identify the direct answer, author, sources, reviewed state, original contribution, and limitations.
  6. Follow the next step on a phone and trigger one harmless validation error.
  7. Write down what is implied but never stated and what is claimed but never evidenced.

Fix those gaps before buying a new acronym.

The durable part of AI-search preparation is straightforward: make the site eligible to retrieve, explicit enough to understand, useful enough to cite, independently credible, and clear enough for a person to use. The aggressive advantage comes from publishing evidence and methods competitors cannot cheaply reproduce—not from publishing the most pages.

Sources