Search is no longer only ten blue links. A growing share of “how do I”, “what is” and “which app” questions gets answered by an assistant that quotes two or three sources. If your site is not one of them, it is invisible to that visitor. Here is the checklist we applied to a small product site, and what we would keep.

1. Write the answer first

Assistants extract. A page that opens with a paragraph of atmosphere and gets to the point in the third section is a page that gets skipped. Each page should answer its own question in the first two sentences: what the thing is, who it is for, what it does. Then elaborate.

2. Make the FAQ real, and mirror it exactly

A FAQPage block in JSON-LD is only useful if the questions and answers are identical to what a visitor can see on the page. Search engines check. We keep the visible FAQ and the structured data as one source, rendered twice, so they cannot drift.

3. Structured data for what the page is about

  • Organization and WebSite on the home page.
  • SoftwareApplication (or MobileApplication) on a product page, with install links to the stores.
  • BlogPosting plus BreadcrumbList on every article.

Leave out what you cannot back up. Aggregate ratings you did not collect and prices that vary by store and country are worse than nothing: they get you flagged, and an assistant that notices inconsistency stops trusting the page.

4. Let the crawlers in, deliberately

The AI crawlers have names: GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Google-Extended and a dozen more. Decide on purpose whether to allow them, then say so explicitly in robots.txt. A blanket allow is fine; silence leaves the decision to defaults you did not choose.

5. Publish a map for machines

A llms.txt file at the root, in the emerging convention: a one-paragraph summary, a handful of key facts, and a list of the pages that matter with a line about each. Generate it from the same content that builds the site so the facts never go stale.

6. The boring basics still count

Canonical URLs, one <title> per page, a description under 160 characters, Open Graph and Twitter tags, a sitemap with real lastmod dates. Assistants lean on the same signals classic search does. A site with duplicate URLs or missing canonicals gets a fraction of its pages indexed, and only indexed pages get cited.

7. Tell the engines when something changed

IndexNow is a single POST with a list of URLs. Bing (and the engines that share its index) recrawl within hours instead of weeks. We ping it from the deploy pipeline after every release; there is no reason to do it by hand.

How to know if it worked

Bing Webmaster Tools has a report for citations in Copilot and its partners, with the queries that grounded them. Google Search Console shows what got indexed and why the rest did not. Both are free, and both are the closest thing to a scoreboard this game has today.

Our own result on a five-page product site: indexed by Bing the same day as the first IndexNow ping, and the pages Google had marked “discovered, not indexed” cleared once we removed a duplicate URL that every sub-page linked to. Small, mechanical fixes. That is most of the job.