Get fetchable, then get quotable. Answer engines pull passages, not pages. Allow OAI-SearchBot and PerplexityBot, stay indexed and snippet-eligible in Google, put the direct answer in the first two sentences under every heading, write paragraphs that stand alone when lifted out, and publish something only you could have written.
Why answer engines behave differently from a results page is covered in the companion piece on AEO versus SEO. This one is the doing.
One uncomfortable fact first. Between them, OpenAI, Perplexity and Google have published almost nothing about how sources get chosen. Every specific tactic circulating is inference from watching outputs. That does not make them wrong. It does mean anyone calling them ranking factors is overselling.
What the platforms actually document
OpenAI’s entire public statement on ordering is that ChatGPT “ranks search results using multiple factors intended to help users find relevant, reliable information.” That is all of it.
Eligibility is concrete, and it is the part worth acting on. OpenAI says any public website can appear in ChatGPT search, and its crawler documentation is blunt: “Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers.” Perplexity documents PerplexityBot as the crawler that surfaces and links sites in its results, separate from Perplexity-User, which fetches on demand and “generally ignores robots.txt rules.” So a page blocked to PerplexityBot still loads when someone pastes the URL, while sitting outside the index answers come from.
Google’s bar is that a page be indexed and eligible to be shown with a snippet, with “no additional technical requirements.” The full crawler and snippet-control list is in the pillar post. Clear this gate first; the rest is wasted on a page nothing can fetch.
Write passages that survive being lifted out
A retrieved paragraph reaches the model with no page around it: no heading above, no preceding sentence, no table it referred to. If it only makes sense in context, it gets dropped or summarized wrong.
- Name the subject in every paragraph. “It typically takes three weeks” is useless alone. “A Google Ads account rebuild typically takes three weeks” travels.
- Delete every internal reference. “As mentioned above,” “see the table below” — each one breaks on extraction.
- Keep the qualifier in the sentence that carries the claim. A number in one paragraph and its condition in the next gets quoted without the condition.
- Put units and time frames inline. “About 40” is not retrievable; “about 40 hours in the first month” is.
One caution, because this gets overdone. Google states that “there’s no requirement to break your content into tiny pieces for AI to better understand it.” The goal is self-contained prose, not a page shredded into fragments.
Answer first, then the nuance
Phrase the heading as the question a buyer would type. Answer it in the next 40 to 60 words. Then the conditions and detail. Most writers do the reverse, and the answer never gets found.
This is the structure that earns featured snippets, and Google is clear you cannot request one: “Google systems determine whether a page would make a good featured snippet for a user’s search request, and if so, elevates it.” Structure is a bid, not a control.
One question per heading. A heading covering three answers none of them cleanly.
Make the entity unambiguous
Before a model names a company it has to be sure which company it is. One business spelled three ways across its own web presence produces answers that name none of them.
Google’s Organization structured data documentation recommends your business name, an indication of real-world presence such as address or telephone, an indication of online presence such as url or logo, and sameAs — the URLs of your profiles elsewhere. Google says to put it on the home page or the page describing the organization, not every page.
Then make the rest of the web agree. Identical name string on the site, the Google Business Profile, LinkedIn, the directories. Named authors with real bio pages rather than “admin.” Be honest about what this buys. Google states that structured data is not required for generative AI search and there is no special schema.org markup you need to add. It is disambiguation insurance, not a citation lever.
Original data is the only durable advantage
Retrieval systems have an unlimited supply of summaries. What they lack is information that exists in exactly one place.
Google’s content guidance asks whether content provides “original information, reporting, research, or analysis,” and whether work drawing on other sources “avoids simply copying or rewriting those sources.” Google has also begun labeling articles that many other stories cite, so readers can find the primary reporting.
For a small business this does not mean commissioning a study. It means publishing numbers you already own: price ranges from your quotes, timelines from jobs you delivered, patterns across the accounts you run. Publish the method and sample size, or it reads as invented.
Dates, done honestly
Nobody documents a recency weighting you can act on. What is documented is how to make your date legible, which matters because OpenAI tells users to check publication dates on cited sources.
Google’s guidance on dates is specific: add a user-visible date and feature it prominently, label it clearly with wording like “Published” or “Last updated,” use datePublished and dateModified in structured data, and make the visible and marked-up dates match, with no future dates. On sitemaps, Google says it uses the lastmod value “if it’s consistently and verifiably accurate,” and that a significant update means the main content, structured data or links — not the copyright year.
Which rules out the tactic everyone reaches for. Bumping dateModified on 200 unchanged pages is not a freshness strategy; it teaches a system that checks your claims to stop believing your dates.
What does nothing
llms.txt
Google states you “don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search.” Neither OpenAI’s nor Perplexity’s crawler documentation mentions the file at all. Publish it if you like; it is not a project. The traffic data is in the pillar post.
Keyword stuffing and variation coverage
Google defines keyword stuffing as “filling a web page with keywords or numbers in an attempt to manipulate rankings.” Its AI guidance goes further: “You don’t have to worry that you don’t have enough ‘long-tail’ keywords or haven’t captured every variation.” Repeating a phrase eleven times does not make a passage more retrievable. It makes it worse prose, which makes it less quotable.
“AI-optimized” content at volume
Tools that spin 300 pages from a keyword list produce what Google’s spam policies name as scaled content abuse: many pages generated primarily to manipulate rankings rather than help users, explicitly including “using generative AI tools or other similar tools to generate many pages without adding value for users.” Past the policy risk, the problem is simpler: volume is the one thing a retrieval system already has too much of.
The order to work in
- Clear the fetch gate. OAI-SearchBot and PerplexityBot allowed, page indexed, no noindex, no restrictive max-snippet.
- Rewrite the opening 60 words so they answer the page’s question outright.
- Re-cut the headings into single questions phrased the way buyers ask them.
- Read every paragraph in isolation. Name orphaned subjects, delete internal references, pull qualifiers into the sentence that carries the claim.
- Add Organization markup on the home or about page, with sameAs pointing at real profiles.
- Fix the name string everywhere it appears off-site; give authors real bio pages.
- Put a labeled, visible date on the page, matched to datePublished and dateModified.
- Find or create the original part. If the page holds nothing you had to run a business to know, it loses to the pages it summarizes.
Steps one through four are a few hours per page. Step eight takes months, and no competitor can copy it in an afternoon.
Where this stops working
Two honest limits. Outputs are not deterministic, so the same question asked twice can cite different sources, and neither run tells you whether your changes worked; the measurement problem is covered in the pillar post. And on queries with an obvious institutional answer, a small business page will not displace a regulator or a manufacturer.
What is winnable is the long middle: specific commercial questions whose best available answer today is a thin roundup written by someone who has never done the work.