A growing share of answers now come from assistants that read a handful of pages and cite the ones they used. Generative engine optimisation is how you become one of them. Guides, an engine-by-engine breakdown, and a free tool, below.
Someone asks how do I fix INP caused by third-party scripts?
The assistant answers
Start by attributing the delay rather than guessing: INP is measured per interaction, so the fix is to find which handler is blocking the main thread. Deferring tag managers and lazy-loading chat widgets until first interaction usually clears the worst of it, though field data takes 28 days to catch up.
And credits these sources
A page does not go from published to quoted in one step. It clears four separate gates, and a page can pass three and still be invisible. Knowing which gate you are stuck at is most of the work, because each one fails for a different reason and has a different fix. Little of this is new: our timeline of how SEO evolved traces the same through-line from keyword matching to machine-readable discoverability.
An AI crawler can fetch the page and read its content without executing your JavaScript.
Stops here if: robots.txt blocks the agent, or the content only exists after render.
The page is pulled into the candidate set for a specific question, not just stored somewhere.
Stops here if: nothing on the page matches how people actually phrase the question.
A specific passage is good enough to quote, so the engine attaches your link to the answer.
Stops here if: the answer is buried under preamble instead of stated in one liftable paragraph.
The engine recognises you as an entity and mentions the brand, not just the URL.
Stops here if: nothing off-site corroborates who you are or what you do.
"AI search" is not one system. Each engine sources its candidates differently, which changes what actually moves your visibility on it. Optimising for all four at once mostly means fixing the foundation they share.
| Engine | Where candidates come from | Links the source | The lever that moves it |
|---|---|---|---|
| ChatGPT | Its own crawler layer for live search, plus content partnerships. Not a straight resale of any one search index. | Yes | Let OAI-SearchBot through, and write answers that stand alone as a single quotable paragraph. |
| Perplexity | Live retrieval at query time, weighted heavily toward recent and specific pages. | Yes | Freshness and specificity. Dated, narrow pages beat broad evergreen ones here. |
| Google AI Overviews | Google's own index. If you are not indexed, you cannot appear. | Yes | Passage-level clarity on pages that already rank. This is the one where classic SEO still does most of the work. |
| Gemini | Google's index plus its entity graph, so brand recognition carries more weight than on the others. | Sometimes | Entity consistency: schema, an About page, and third-party mentions that agree with each other. |
One caveat worth stating plainly: none of these engines publishes its ranking logic, and all four change quietly. Treat the right-hand column as where the evidence currently points, not as documented behaviour. Our study of what AI Overview traffic is actually worth covers what happens after the citation lands.
Four claims we hear on nearly every call, and what the evidence actually supports.
Retrieval happens per question, not per domain. An engine assembling an answer needs a passage that resolves that specific question, and a narrow page from a small site often does that better than a broad page from a large one. Authority still helps at the margin, but it is not the gate it is in classic rankings.
No engine has confirmed it uses llms.txt for retrieval or ranking. It is a proposal, and a reasonable one, but shipping it does not buy visibility. What it actually does today is worth ten minutes of reading before you spend a day on it.
It protects it from being read, which also means it protects it from being cited. Training crawlers and search crawlers are often separate agents, so the useful decision is per agent, not a blanket block. Decide what you are trading away before you write the disallow line.
Google's stated position targets low-value content produced at scale, not the tool used to write it. The failure mode is thin, undifferentiated pages, which was already a failure mode. We unpacked the actual policy and what it does and does not cover in does Google penalize AI-generated content.
Each guide covers one decision end to end, with the implementation detail rather than the summary.
The full playbook: which crawlers to let through, how to shape a passage so it can be lifted whole, the entity work that gets you named rather than just linked, and how to check whether any of it worked. This is the one to read first if you only read one.
Read the playbookWhat the file is, which engines have committed to reading it, and an honest answer on whether it is worth shipping yet.
Read the guideBuild the segment that separates assistant referrals from the rest, so you can prove the channel exists before you invest in it.
Read the guideWhat the spam policy actually says, where the line sits in practice, and the editorial process that keeps you on the right side of it.
Read the guidePoint it at your site and get a structured llms.txt back, ready to review and upload. No signup.
Open the toolA plain-text map of what your site covers and where the canonical version of each thing lives. Nothing here requires a signup, and the output is yours to edit before it goes anywhere.
GEO is not a project that finishes. Engines change their retrieval quietly and without notice, so the work is a loop you re-run, not a checklist you complete.
Establish what is already being cited, and by which engine, before changing anything. Without a baseline you cannot tell improvement from noise, and a migration can quietly cost you the lot.
Fix the machine-readable layer: crawler access, entity markup, and internal linking that makes the relationships explicit. Most of this is schema work.
Rewrite the pages that should be cited so each one answers its question in a single liftable paragraph, near the top, in the phrasing people actually use.
Track assistant referrals as their own channel, using the GA4 setup. Expect Search Console and GA4 to disagree on the numbers.
It shares about 80% of its foundation with classic SEO: crawlability, structure, and content that genuinely answers something. The remaining 20% is a real difference in target. Classic SEO optimises a page to outrank other pages for a query. GEO optimises a passage to be extracted and quoted inside someone else's answer. That changes how you write, where you put the answer on the page, and what success looks like, because a citation can happen without a click.
No major engine has confirmed that it reads llms.txt for retrieval or ranking. It is a community proposal, not a standard any of them have adopted. That said, it costs very little to publish, it forces a useful audit of what your site actually covers, and it is cheap insurance if adoption arrives. Ship it if it takes an hour. Do not restructure your content around it.
Two ways, and you want both. Ask the engine directly: run the twenty questions your buyers actually ask and record which sources come back, repeated monthly so you have a trend rather than a snapshot. Then confirm it in your analytics by segmenting referral traffic from assistant hostnames, which tells you whether those citations are sending real people. Citations without referrals still have value as brand exposure, but you should know which one you are getting.
Only if you have decided you would rather not be cited. Training crawlers and search-retrieval crawlers are usually separate user agents, so you can block one and allow the other. Blocking the retrieval agent removes you from the candidate set entirely, which is a visibility decision dressed up as a security one. Publishers with licensing leverage sometimes take that trade deliberately. Most businesses should not.
You do not need it to be cited, and plenty of cited pages have none. It helps with a different job: being recognised as an entity rather than as an anonymous URL, which is what gets your brand named in an answer instead of just linked. Organization, Person, and Product markup that stays consistent with the rest of the web is the part that matters. Marking up an FAQ block will not get you retrieved.
Much less than it decides rankings. Retrieval is scoped to a specific question, and the winning passage is the one that resolves it most directly. That is why niche sites routinely appear alongside major publishers in the same answer. Authority still helps as a tiebreaker and in how much an engine trusts a contested claim, but a well-shaped page on a small domain is genuinely competitive here in a way it is not on page one.
It depends on whether the engine retrieves live or from an index. Engines that fetch at query time can surface a page within days of publication. Engines working from a search index inherit that index's crawl schedule, so you are waiting on normal indexation first, which can be days for a well-linked site or weeks for a new one. Either way, judge the result over a month of repeated checks, not on the first attempt.
Traditional search and generative assistants reward the same technical foundation. We build it: machine-readable structure, crawl paths that do not depend on JavaScript, and content shaped to be quoted. Start with the free tool, or bring us the questions you want to be the answer to.
No signup for the tool · Works on any CMS · UK-based, direct with the person doing the work