LLM SEO: The 2026 Playbook for Getting Cited by AI

Most of the LLM SEO advice published this year asks you to rebuild your site for machines. Chunk every page. Ship an llms.txt. Invent a schema for answer engines. I did the boring version instead and read what the engines themselves publish. The honest summary: the rewriting is mostly optional, the plumbing is not.

Google says so out loud. Its guide to optimizing for generative AI features states that "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search," that "There's no requirement to break your content into tiny pieces for AI to better understand it," and that "Structured data isn't required for generative AI search." The eligibility bar is older than any of this: "a page must be indexed and eligible to be shown in Google Search with a snippet."

So the job is not a new content genre. It is four unglamorous tasks: let the right crawlers in, exist in the index each engine reads, write something worth lifting, and measure whether it landed.

What LLM SEO actually is, once you strip out the folklore

Answer engines do not remember your website. They retrieve it. ChatGPT's search answers come from pages a search crawler collected, Perplexity's from its own crawl, AI Overviews from the same Google index that serves the blue links. If the retrieval layer cannot reach you, the model has nothing to cite however good the prose is. Every engine therefore hands you an explicit on/off switch, and most sites have never looked at theirs.

Step 1: check which bots you actually let in

The vendors publish the user-agent tokens, so this is checkable rather than debatable. OpenAI runs four crawlers, and one of them decides whether you show up in ChatGPT search. OpenAI's crawler documentation says "Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers." GPTBot is the training crawler, and ChatGPT-User is the live fetch, which OpenAI notes "is not used to determine whether content may appear in Search."

Anthropic splits the same way. Its crawler page lists ClaudeBot for training, Claude-User for live questions and Claude-SearchBot for search indexing, warning that disabling the last one "may reduce your site's visibility and accuracy in user search results."

Perplexity keeps it to two. PerplexityBot is "designed to surface and link websites in search results on Perplexity. It is not used to crawl content for AI foundation models." Perplexity-User is the live fetch, and the docs admit it "generally ignores robots.txt rules."

The useful consequence: refusing to feed model training and staying citable are compatible positions. Block GPTBot and ClaudeBot, allow OAI-SearchBot, Claude-SearchBot and PerplexityBot, and you keep the citations without donating the corpus. My own robots.txt is a plain Allow: / with the API path closed, because I decided the training exposure was worth the reach. Make it a decision, not a default nobody has read since launch.

Step 2: exist in the index the answer is built from

Being allowed is not the same as being present. Google's bar, from the guide above, is the ordinary one: indexed, snippet-eligible, technically sound.

Bing deserves more attention than its search share suggests: Microsoft now reports AI citations there. The AI Performance report in Bing Webmaster Tools, in public preview, covers "Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations." If your site has never been verified there, do that before writing another paragraph of content.

And yes, I publish an llms.txt. It is cheap to maintain and it is not a strategy: Google's guide says Search ignores such files.

Step 3: write the paragraph a model can lift

The content half of Google's guide is unusually direct: it asks for "non-commodity content" carrying "unique expert or experienced takes that go beyond common knowledge." Translated for a founder writing their own site, another rephrasing of the top three results is invisible, because the model already has that summary.

What travels well in practice:

• Answer the question in the first two sentences, before any context-setting.
• One claim per paragraph, with the number and its basis in the same sentence.
• Your own comparison table, so there is a structured object to lift rather than a paragraph to interpret.
• Specifics only you have: your prices, your test, your failure, your dates.

The engine-specific version of this is in how to rank in ChatGPT.

Step 4: measure your LLM SEO, or you are just guessing

Two dashboards exist now and almost nobody I talk to has opened either. Google's guide points to the Generative AI performance report in Search Console to "measure how your content is performing in generative AI features on Google Search." Bing's AI Performance report gives total citations, average cited pages per day, page-level citation counts, and the one column I find genuinely new: grounding queries, described as "key phrases the AI used when retrieving content that was referenced in AI-generated answers." Those are phrases the machine invented on its way to your page, not phrases a human typed. A keyword list you could not have bought.

Neither dashboard tells you what a rival gets recommended for. That is why the other half of measurement is asking the engines your buyers' questions yourself and logging who got cited.

The clicks are not coming back the way they left

Set expectations first. Cloudflare tracks a crawl-to-refer ratio: HTML page requests from an AI platform's crawlers divided by the referrals that platform sends back. In the week of 19 to 26 June 2025 the spread ran from Anthropic at roughly 70,900 crawl requests per referral down to Mistral at 0.1:1, ten referrals per crawl. Cloudflare notes the ratios shift week to week, so take the shape rather than the digits: reading has decoupled from clicking.

If sessions are your only KPI, this work looks like a failure while your brand is being recommended in rooms you cannot see. Track citations and mentions first; treat referral traffic as a bonus.

Make it a weekly loop, not a project

None of the four steps is hard. They recur, and recurring work is what a solo founder drops by week three.

So I automated it. A scheduled Notis automation holds about twenty buyer questions, asks each one, reads the pages that come back, records which domains got cited into a Notis database, and sends me the week-over-week diff as a message on Monday morning. I never open a dashboard; the dashboard opens me.

The economics are boringly small. Reading a web page through the on-demand catalogue costs about $0.004, so a twenty-question run is a few cents. Pro is $13 a month billed annually and includes $20 of usage a month, which a weekly visibility run barely dents, and on-demand usage past the allowance keeps working and is billed in arrears. There is a hard stop underneath it: any single pay-per-use call quoted above $1.20 is refused, and the agent tells me the price and asks before retrying.

Wiring it yourself is fair too, and worth pricing honestly: on the Claude API, web search is $10 per 1,000 searches on top of tokens, and Claude Sonnet 5 runs $2 per million input tokens and $10 per million output. What you buy by not building it is the delivery: the log landing in Notion and the recap arriving in a chat thread you already read, instead of a cron job you maintain.

What I would do in the next two weeks

• Never touched robots.txt. Open it, add explicit rules for OAI-SearchBot, Claude-SearchBot and PerplexityBot, then decide deliberately about GPTBot and ClaudeBot. Half an hour, once.
• Already ranking on Google. Skip the content panic. Verify Bing Webmaster Tools, open both AI reports, mine grounding queries for questions you have not answered.
• Pre-traffic. Indexation first. An engine cannot cite a page it never retrieved.
• Everyone. Write down twenty questions a buyer would type, check them monthly, log who gets cited. That log is your scoreboard.

The glamorous half of this work, rewriting everything for robots, is the half the engines say they do not need. The unglamorous half, permissions and measurement, is the half nobody runs. That asymmetry is the opportunity, and it will not last.