Generative engine optimization is the work of being useful to an AI answer rather than only to a ranked list. Most of what is written about it is confident and unverifiable. This guide separates what is actually known from what is being guessed, and ends with a checklist you can run against your own pages today.
Generative engine optimization (GEO) is making a page likely to be used, and ideally cited, when a generative system answers a question. Answer engine optimization (AEO) is used for the same idea, usually with more emphasis on direct answers and featured-snippet-style formatting. LLM SEO is a looser label for the whole area and occasionally for something different — getting a model to say nice things about your brand from its training data, which is not a thing you can optimise on a deadline.
The distinctions matter less than one structural fact: most AI answers about current information are retrieval-augmented. The system runs a search, reads some pages and writes a summary with links. The search part is a search index. That is why the dull half of SEO — crawlable, indexable, canonical, fast, structured — did not stop mattering when AI answers arrived. It became the entry fee.
Worth separating, because the confident half of the internet is mostly in the right-hand column.
1. Be fetchable. The most common reason a page is not cited is that it was not retrieved. Check robots.txt, check that the page is not noindex by accident, check that it is in your sitemap and that the sitemap returns 200.
2. Answer in the first paragraph. If the page exists to answer a question, answer it before the context, the history and the pitch. A passage that needs three paragraphs of setup cannot be lifted.
3. Use headings as questions where that is honest. Not keyword-stuffed H2s — real questions people ask, with the answer directly underneath. Your own Search Console query list is the best source for the phrasing, because those are questions you are already being shown for.
4. Make claims checkable. Say where a number came from and when. This is good practice generally and it happens to be what makes a passage safe for a system to quote.
5. Declare what the page is. Valid JSON-LD for the page type, the author and the date. Not because schema forces a citation, but because a broken or missing declaration leaves it to be inferred.
6. Kill the near-duplicates. Forty pages that differ by a town name are the least quotable content you can publish, and they are a real risk of being treated as low quality in ordinary search too.
7. Then stop. The rest of the advice circulating is untested. Spending a quarter on it instead of on the seven items above is a poor trade.
MrKeyword's crawler runs 60 technical checks. These are the ones that matter for being retrieved and quoted — named by their real ids so you can match them to what the audit reports.
| Check | What it finds | Why it matters here |
|---|---|---|
invalid_structured_data | Invalid structured data (JSON-LD) | Broken markup is worse than none: it is parsed, rejected, and you never find out. |
missing_structured_data | Pages without structured data | Schema does not force a citation, but it is how you state author, date and type unambiguously. |
missing_h1 | Missing H1 | A passage is easier to lift when the page says what it is about in one heading. |
multiple_h1 | Multiple H1 tags | Several competing H1s make the topic of a page ambiguous to anything parsing it. |
thin_content | Thin content | There has to be an answer on the page before anything can quote it. |
duplicate_content | Duplicate content | Two pages making the same claim split the evidence and neither becomes the source. |
near_duplicate_template | Near-duplicate template pages | Programmatic pages that differ by a city name are the least quotable content on the web. |
missing_canonical | Missing canonical tag | Without one you are asking a retrieval system to pick which URL is the real answer. |
blocked_by_robots | Linked URLs blocked by robots.txt | The most common cause of not being cited is not being fetchable. |
missing_meta_description | Missing meta description | A written summary is the version of your page that gets shown when something else writes the snippet. |
missing_lang | Missing page language | Cheap to fix, and it tells every parser which language the answer is in. |
Every one of these is a rule, not a model: the same page gives the same answer every run. How the site audit works.
It does not track AI citations. MrKeyword cannot tell you whether you appeared in an AI Overview, whether ChatGPT cited you, or what your share of AI answers is. Collecting that honestly at scale is not something we do today, and a number we cannot show the working for is worth less than no number.
It does give you the groundwork. The technical checks above, run over your whole site, with the fix written out for each and sorted by how many pages each one affects.
It does show you the question queries you already win. Search Console knows every query your site has been shown for. Filtering those for question phrasings is the most grounded input to GEO work that exists, because it is a record of real people asking real things and seeing your page.
And it will draft the page. AI content briefs turn a target page, its audit and its real query data into an outline, the questions to answer and the keywords to cover, each tagged with the evidence behind it. Briefs are on Pro AI, or via the 50 AI page audits pack.
GEO is where the industry's claims are loosest, so here is ours in plain terms.
Generative engine optimization, or GEO, is the practice of making a page likely to be used and cited when an AI system answers a question — Google's AI Overviews, ChatGPT with search, Perplexity, Copilot. Answer engine optimization (AEO) is used for roughly the same idea. It is not a separate discipline with its own secret levers; it is mostly ordinary technical and content SEO judged by a different reader.
Less than the amount written about it suggests. The retrieval layer behind most AI answers is a search index, so being crawlable, indexable and clearly structured still decides whether you are a candidate. What changes is the unit: a classic result rewards a page that ranks, a generative answer rewards a passage that can be lifted and attributed. That favours direct answers, clean headings and claims that are easy to verify.
No. MrKeyword does not measure whether you are cited in AI Overviews, ChatGPT, Perplexity or Copilot, and we are not going to sell a number we cannot collect honestly. What it gives you is the technical groundwork and your real Search Console data — including which question-style queries you already appear for.
Not one that means anything. Nobody outside those systems can see how a passage was selected, and any single score claiming to predict it is a model built on guesses. The checklist on this page is made of things that are verifiable on your own page.
It is a proposal, not a standard, and no major search or AI system has committed to reading it. It costs almost nothing to publish, and it does not replace being crawlable. Treat it as a bet, not a task.
That is a business decision, not an SEO one. Blocking GPTBot or similar removes you from some training and retrieval pipelines, which protects content and also removes a route to being cited. MrKeyword's crawler reports what your robots.txt says so you can see the decision you have actually made, rather than the one you think you made.
Publish a direct answer near the top of the pages that deserve one, fix the structured-data and heading issues the crawler flags, and look at your Search Console queries for question phrasings you already appear for. Those three are real work with real evidence behind them, which is more than most GEO advice offers.
All 60 checks, in three severities, with the fix for each.
Read moreFind the question phrasings people actually search for.
Read moreAsk your assistant which question queries you already appear for.
Read moreRun the technical half of this checklist on your own site for nothing.
Read moreIt is the part that is actually known to matter, and the Free plan covers 100 pages a month.