Most GEO checklists mix three very different things: findings from the one substantial piece of published research, behaviour the platforms document themselves, and the author's opinion. Every item below says which it is, so you can weight them yourself.
The order is the opinionated part. Reachability comes before structure, structure before content, content before authority — because an engine cannot cite a page it cannot fetch, and selling content work to a site in that state is selling a lever that cannot move.
1. Reachability
An engine cannot cite a page it cannot fetch. Everything else is wasted until this is true.
Explicitly allow the AI crawlers in robots.txt
Platform-documentedGPTBot, ClaudeBot, PerplexityBot, Google-Extended, OAI-SearchBot and Applebot-Extended each read robots.txt. Cloudflare began blocking several by default in 2024, so a site can be blocking them without anyone having decided to.
Verify: curl -A "GPTBot" https://yoursite.com/ and confirm a 200, then check robots.txt lists each agent.
Serve meaningful HTML without JavaScript execution
Platform-documentedMost AI crawlers do not run JavaScript. A client-rendered page returns an empty shell to them regardless of how it looks in a browser.
Verify: curl the URL and confirm the H1 and body copy are present in the raw response.
Fix internal links that 404 or pass through a redirect
Our operating experienceBroken navigation spends crawl budget on nothing. We found 35 broken internal links on our own site, including three in the main navbar that appeared on every page.
Verify: Crawl every internal href and assert a direct 200 — no 301/308 hops.
Give every page you want cited at least one contextual internal link
Our operating experienceSitemap-only discovery is a weak signal. 22 of our 27 industry pages had zero contextual in-links.
Verify: Count in-links per URL from rendered HTML, excluding nav and footer chrome.
2. Distinctness
Interchangeable pages give an engine no reason to prefer any one of them.
Give every page a unique meta description
Our operating experienceDuplicates signal that pages are substitutes. A prerender fallback on our site silently overwrote 108 pages with one identical description.
Verify: Extract every description and assert the count of unique values equals the page count.
Keep titles unique and under about 60 characters
Our operating experienceTruncation removes the part that identifies the page. Ours ran to 185 characters, 42 of which were a repeated site-name suffix.
Verify: Assert no duplicate titles and flag any over 60 characters.
Use absolute self-referencing canonicals
Platform-documentedA relative canonical is invalid, and a canonical pointing at a different URL tells the engine to cite that one instead.
Verify: Assert every canonical starts with https:// and matches its own URL unless deliberately consolidating.
3. Extractability
Make the quotable sentence easy to find. This is the part the research is clearest about.
Open with a direct, self-contained answer
Peer-reviewedA model lifts passages that survive removal from context. A page that builds to its conclusion gives it nothing clean to quote.
Verify: Read the first 60 words alone. If they do not answer the title, rewrite them.
Add a key-takeaways block of complete declarative sentences
Our operating experienceEvery competitor page currently winning AI/affiliate queries opens with one, including pages far shorter than ours that outranked us.
Verify: Confirm the block renders in the raw HTML, not only after hydration.
Include specific numbers, ranges and dates
Peer-reviewedThe KDD 2024 GEO study found that adding statistics, quotations and citations raised a source’s visibility in generated answers by up to 40%.
Verify: Count concrete figures per page; a page of adjectives has nothing to extract.
Cite and link primary sources by name
Peer-reviewedCitation was among the strongest levers measured in the same study. An unlinked "research shows" is weaker than a named, linked paper.
Verify: Every claimed statistic resolves to a source a reader can open.
Do not keyword-stuff
Peer-reviewedThe same study found keyword stuffing measurably reduced visibility. It is one of the few tactics with evidence of harm rather than mere ineffectiveness.
Verify: Keep target-phrase density natural; if a sentence exists to hold a keyword, cut it.
4. Machine legibility
Remove ambiguity about what the page asserts and who is asserting it.
Add Article, FAQPage and Organization structured data
Platform-documentedStructured data is not an AI ranking factor, but it removes guesswork about the page’s claims, author and subject.
Verify: Validate the JSON-LD and confirm it matches the visible content rather than contradicting it.
Make Organization sameAs point only at profiles that resolve
Our operating experienceA sameAs pointing at a 404 weakens entity resolution instead of supporting it. Ours listed a LinkedIn company page and an X handle that both 404’d.
Verify: Request each sameAs URL and assert a 200.
Publish an llms.txt
Platform-documentedNo major engine is known to use it as a ranking input today. It is cheap and makes site structure explicit — treat claims of measured lift from llms.txt alone with scepticism.
Verify: Fetch /llms.txt and confirm it lists your real sections, not a stale copy.
5. Measurement
Without a baseline, any later claim of improvement is unfalsifiable.
Record a citation baseline before changing anything
Our operating experienceCitation behaviour also shifts when models are retrained, so a single post-hoc check cannot distinguish your work from model drift.
Verify: Fix a prompt set of 15–25 real buying questions, run them across engines, and log whether you are named and where.
Keep control prompts you do not expect to win
Our operating experienceIf both your target and control prompts improve, the change was probably the model, not you.
Verify: Split the prompt set into target and control groups and track them separately.
Re-run weekly, not once
Our operating experienceAnswers vary between runs for the same prompt. One observation is an anecdote.
Verify: Append dated rows rather than overwriting; compare trend, not single readings.
Frequently asked questions
How do I get cited by ChatGPT?
Make the page fetchable by GPTBot and readable without JavaScript, give it a distinct title and description, open with a direct self-contained answer, include specific figures, and cite primary sources by name. The KDD 2024 GEO study found citations, quotations and statistics raised source visibility by up to 40% while keyword stuffing reduced it. There is no way to purchase placement: OpenAI states product results are organic and unsponsored.
Is there a GEO checklist that actually has evidence behind it?
Partly. The peer-reviewed base is essentially one paper — Aggarwal et al., "GEO: Generative Engine Optimization" (KDD 2024) — which supports answer-first structure, statistics, quotations and citations, and finds keyword stuffing harmful. Crawlability, unique metadata and working internal links are standard technical SEO whose value predates AI search. Anything beyond those two categories is currently practitioner opinion, including ours.
Does llms.txt improve AI visibility?
There is no public evidence that a major AI engine uses llms.txt as a ranking input today. It costs almost nothing and documents your structure clearly, which is why we publish one, but treat any vendor claiming measured lift from llms.txt alone with scepticism.
How long does GEO work take to show results?
Structural changes are picked up on normal re-crawl cadence, so weeks rather than days. Citation behaviour also changes when models are updated, which is why a dated baseline and weekly re-runs matter more than a single before-and-after check.
What should I fix first?
Reachability. An engine cannot cite a page it cannot fetch, so broken internal links, JavaScript-only rendering and blocked crawlers outrank content work. Only after that do distinctness, extractability and structured data pay off.
Related: what we changed on our own site, what GEO actually costs, and the AI marketing hub.