How to Rank in AI Search Engines at the Page Level: ChatGPT, Perplexity, Gemini

Quick answer

To rank in AI search engines, Canlah AI runs six page-level steps in a fixed order, and the first three decide whether an engine can copy the page at all: allow the retrieval crawler, serve the answer as text in the HTML and open each section with a standalone answer that carries the number. In a Canlah AI sample of 526 citations from ChatGPT and Google AI Mode across 60 buyer questions and 92 cited pages, roughly 73% of the copied facts sat in the first 30% of the page, and the copied sentence was almost always the first sentence of a paragraph. Google sets the technical bar for its AI features at the Search level: a page must be indexed and eligible to be shown in Google Search with a snippet, with no additional technical requirements.

Canlah AI publishes here the manual version of the on-site stage it runs for clients, and its own site shows the limit of that stage. On September 7, 2026 every page-level check on canlah.ai passed, yet the site was named in 0 of 12 category-level ChatGPT answers and scored 4 of 100 on AI citation.

Treat page-level work as necessary and not sufficient. The six steps below change what an engine can copy from a page; none change what other websites say about you.

Disclosure: Canlah AI publishes this guide and sells the managed version of the work described below. Google, OpenAI and Perplexity documentation was read on current public pages on September 23, 2026 and was not tested through paid accounts. The Canlah AI figures come from a self-audit of canlah.ai run on September 7, 2026 and are a single-date sample.

What does ranking in AI search engines actually mean?

Ranking in AI search engines means four measurable outcomes rather than a rank position, and a page-level fix moves only the first two.

• Retrieval: The page was fetched or indexed and the engine’s retrieval crawler reached it.
• Citation: One of your URLs sits among the supporting links, and it supports an identifiable sentence of the answer.
• Mention: The answer names your brand, product or price, with or without a link.
• Recommendation: The answer presents your brand as suitable for the stated need rather than as one option of several.
• Question type: The buyer either named your brand or asked an open question such as “best GEO agency in Singapore”.

The last line decides which half of the work matters. A site named on branded questions and absent on open ones has no page-level defect: every engine can already read it, and the gap sits in what other pages say.

Six page-level steps that apply across engines

Six page-level steps apply across every engine surface reviewed, and they run in order because each one either clears a cause or shows that the cause sits off the page.

Step 1: Allow the retrieval crawler, not only the training crawler

Retrieval crawlers and training crawlers are separate tokens in robots.txt, and blocking the wrong one removes you from answers without withholding training data. OpenAI documents OAI-SearchBot for search, stating that sites opted out of it will not be shown in ChatGPT search answers, while GPTBot governs training. Perplexity documents PerplexityBot for search results and Perplexity-User for user-triggered visits. Both state that a robots.txt change takes about 24 hours to reach them.

Check the tokens against the live file, then the layer above it. A CDN rule, a WAF challenge or a rate limit can refuse a crawler that robots.txt allows, which is the failure Canlah AI sees most often on sites that believe they are open to everyone.

Step 2: Serve the answer as text in the HTML the crawler receives

An engine can only copy text it receives, so the answer must survive in the raw HTML rather than appearing after a click, a tab or a script. Google’s guidance asks site owners to make important content available in textual form and states that Google can process content within JavaScript as long as it is not blocked. Read the HTML Googlebot received in URL Inspection, then fetch the page with a plain request to see what a non-rendering crawler gets. Price tables inside images, specifications inside PDFs and answers inside click-loaded accordions are the three failures worth checking first.

Step 3: Make the first sentence of each section a standalone answer

Across the 92 cited pages in the sample, the copied sentence was almost always the first sentence of a paragraph, and in the 18 legal pages analysed it was the paragraph-opening sentence in 15. Rewrite each H2 so its first sentence names the subject, states the answer, carries the number and survives being lifted out of the page. The cleanest counter-example was a page that opened a section with “is a complicated matter”; that sentence was not copied, and the conclusion in its FAQ was.

Step 4: Put every key number in the first 30% and repeat it in a table

Roughly 73% of the copied facts in the sample sat in the first 30% of the page, and tables were the most copied container, header row included. Put the exact figure in the opening block, then repeat it in a table whose headers carry the judgment, such as “Best for”, “Market range” or “What to verify”. Make the summary and the table agree exactly: where the two disagreed, engines used the table and skipped the summary.

Step 5: Keep structured data consistent with the visible text

Structured data is not an entry requirement for AI answers, and Google states there is no special schema.org markup needed to appear in AI Overviews or AI Mode. It still asks that any structured data you publish matches the visible text, which makes an inconsistent block a liability rather than a neutral extra. Publish Organization and Article for rich results, validate them, then check that every price, name and date matches what a reader sees.

Step 6: Describe the entity the same way on every page and off-site

Engines assemble a brand description from many sources, so a category name that changes between the home page, the pricing page and directory listings weakens all three. Write one sentence stating what the company is, for whom, where and one verifiable fact, then repeat it on the site, in the Business Profile, in directories and in structured data. Canlah AI keeps that sentence on a public /ai-info page so the wording an engine finds is the wording the company approved.

The page-level checklist at a glance

Six page-level steps, six free checks and six decisions fit on one screen, and step 1 gates the other five because a blocked crawler makes every later fix unreadable.

Six page-level steps for AI search visibility, the free check for each and the decision it supports. This is the on-site sequence Canlah AI runs before off-site work, reviewed September 23, 2026.

Step	What to change on the page	How to verify it free	Decision it supports
1 Crawler access	Allow OAI-SearchBot, ChatGPT-User, PerplexityBot and Googlebot in robots.txt and at the CDN	Read /robots.txt; check server logs per user agent	Fix access before writing anything
2 Text in HTML	Move answers out of images, PDFs, tabs and click-loaded blocks	URL Inspection; a plain curl of the URL	Fix rendering before rewriting copy
3 Answer-first sentences	Rewrite the first sentence under each H2 as a liftable answer carrying the number	Read the first 30 words under each H2 out of context	Rewrite passages, not the page
4 Numbers up front	Put each key figure in the opening block and in a table with judgment-carrying headers	Count the digits in the first 30% of the page	Restructure the page order
5 Structured data	Publish Organization and Article, match every field to visible text	Rich Results Test; compare each field with the page	Correct markup, do not add types
6 Entity consistency	One approved sentence for what the company is, repeated everywhere	Search the brand name and read the first 10 descriptions	Standardise before off-site outreach

Source: the Canlah AI on-site checklist and the vendor documentation listed under Method and sources, reviewed September 23, 2026. Free tools are named for the check they support, not ranked, and no free checker replaces a panel whose engines, market and run count you control.

Which crawler controls which engine

Four engine surfaces name the robots.txt token that controls their retrieval: ChatGPT search uses OAI-SearchBot, Perplexity uses PerplexityBot, Google AI Overviews and Google AI Mode use Googlebot and Gemini grounding outside Search uses Google-Extended. For DeepSeek, Qwen and Doubao the token was not found on the reviewed developer pages.

Robots.txt tokens for five engine surfaces, what each vendor states a block costs and the check each one needs, from documentation read September 23, 2026.

Engine surface	Robots.txt token to allow	What the vendor states a block costs	What to verify
ChatGPT search	OAI-SearchBot, plus ChatGPT-User for user-triggered fetches	Sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can appear as navigational links	Both tokens in /robots.txt, then the CDN rule above it
Perplexity	PerplexityBot, plus Perplexity-User	PerplexityBot surfaces and links websites in Perplexity search results and is not used to crawl for foundation models	That the search token is allowed separately from the training decision
Google AI Overviews and Google AI Mode	Googlebot	Robots.txt directives for Googlebot are the control; the page must be indexed and snippet-eligible	Indexing plus nosnippet and max-snippet, in URL Inspection
Gemini grounding outside Search	Google-Extended, a product token rather than a separate crawler	It limits training and grounding in some of Google’s other systems and does not affect Search or AI Overviews	That it is decided on training policy, not on AI search visibility
DeepSeek, Qwen and Doubao	Not found on reviewed page	Not found on reviewed page	Allow each market’s main search crawlers, then test one branded question in each assistant

Sources: OpenAI crawlers page, Perplexity crawler documentation and Google AI features documentation, read September 23, 2026. “Not found on reviewed page” means the vendor’s public developer materials did not name a retrieval token or its effect when read on that date. It is not proof that no such crawler exists.

What a 526-citation sample shows about page format

The Canlah AI study behind this guide collected 526 citations from ChatGPT and Google AI Mode across 60 buyer questions in five categories, then read 92 cited pages line by line. Three format findings held in every category.

First, the two engines cite different page types. ChatGPT went to source pages: government pages, official pricing pages and product pages. Google AI Mode went to secondary syntheses: rankings, comparison pages, videos and forum threads. A page that wants both needs two sentences, one exact own-side fact and one market range that names its variables.

Second, position beat polish and length was not a criterion. Cited pages ran from 40 words to about 9,000, median near 1,300 to 1,500, and disorganised pages were cited when they carried an exact figure early.

Third, the FAQ was rarely the copy source: about six of the 75 locatable copied facts came from an FAQ block, and most of those pages never stated the conclusion in the body. Keep the FAQ for schema and never leave a conclusion only there.

What Canlah AI’s own audit showed when every page-level check passed

Canlah AI audited canlah.ai on September 7, 2026 with the pipeline it runs for clients, and the page-level layer came back clean while the AI-citation dimension scored 4 of 100. The crawl reached 80 pages against a 174-URL sitemap and found 0 pages missing a canonical tag, 0 missing a title, 0 missing a meta description and 0 missing an H1, with hreflang configured. Every documented AI retrieval crawler was allowed, including OAI-SearchBot, ChatGPT-User, PerplexityBot, Claude-SearchBot and Claude-User. Structured data parsed without errors across three JSON-LD blocks, scoring 60 of 100 with Organization and Article absent.

The answer layer did not follow. On category-level questions the site was named in 0 of 12 ChatGPT answers and 0 of 12 Gemini answers. On main-shelf questions it reached 3 of 20 and 0 of 24. On branded verification questions it was named in 11 of 11 and 12 of 12.

A page every engine can read is necessary for AI search visibility and does not produce it.

The same run listed the domains cited instead, led by aistudio.com.sg with 91 citations, mediaplus.com.sg with 85 and stridec.com with 78.

What does not move AI search rankings

Google’s guide to optimizing for generative AI search names five practices that its Search systems do not need, and the six steps above skip all five.

• LLMS.txt and other special files: Google states that Search, including its generative AI features, does not use them.
• Chunking content: Google states there is no requirement to split content into small pieces and no ideal page length; cited pages in the Canlah AI sample ran from 40 to 9,000 words.
• Rewriting content for AI systems: Google states that its systems understand synonyms and intent, so one page per phrasing is unnecessary.
• Seeking inauthentic mentions: Google states that its spam systems apply to manufactured mentions.
• Over-focusing on structured data: Google states that no special markup exists for generative AI search.

Google also states that no third-party tool has access to its internal ranking or AI systems, and that the Generative AI performance report in Search Console is the first-party measurement. Treat a vendor claim about an internal Google AI metric as a reason to ask for raw answers.

How to interpret five common results

Five result patterns cover almost every page-level finding, and each points at a different fix rather than at more pages.

Cited but not named. The engine used the page and credited no brand. Check whether the copied passage states the brand name in the same sentence as the fact, not only in the heading above it.

Named on branded questions only. This is the most common pattern and the one Canlah AI found on its own site, at 11 of 11 branded answers against 0 of 12 category answers. Every engine can read the site and none connects it with the category, so the constraint is off-site.

Visible in ChatGPT, absent in Google AI Overviews. The Canlah AI sample showed ChatGPT favouring source pages and Google AI Mode favouring secondary syntheses. Diagnose each engine separately: an exact own-side fact serves ChatGPT, while Google needs the brand on the rankings and comparison pages it already cites.

Named with the wrong number. Correction outranks exposure. Align the page, the structured data, the Business Profile and directories on one approved fact sheet, then re-run the question that produced the error.

Cited in one run and absent in the next. Record frequencies such as “cited in 2 of 6 runs” and never report a single run as a rate.

What page-level work cannot prove

Page-level work cannot prove three things, and each gap is a reason to keep an off-site measurement running beside it.

• It cannot prove that a rewrite caused a later citation, because the index, the models and competitor pages all change in the same window.
• It cannot produce category-level mentions on its own, which is what the Canlah AI result shows.
• It cannot correct a description that third-party pages state wrongly, because the engine reads those pages.

When to move from page-level work to off-site evidence

The move from page-level work to off-site evidence is justified once all six steps pass and open category questions still return other brands, which is itself the diagnosis. The remaining work is earning a place on the pages engines already cite for the category: listicles, directories, reviews, community threads and original research rather than more pages on your own site.

Where Canlah AI fits

Canlah AI is a Singapore-based SEO + GEO agency that measures and improves brand visibility inside ChatGPT, Gemini, Google AI Overviews and Google AI Mode, reporting AI visibility as re-verifiable ranges with timestamped evidence. It sells the page-level stage as a fixed-scope AI-readability rebuild and the off-site stage inside its GEO service, priced per AI platform from $1,500 a month, after a free AI visibility report within 48 hours.

The limitation sits in Canlah AI’s own numbers: a site with complete page-level hygiene scored 4 of 100 on AI citation, so no provider can sell page-level work as a route to category visibility. Ask any provider to re-run your panel in front of you and show the raw answers and cited URLs. Verify the diagnosis against evidence you can inspect, not a composite score.

Frequently asked questions

How do you rank in AI search engines?

You do not rank in an AI answer; you get retrieved and then copied. To rank in AI search engines, allow each engine’s retrieval crawler, serve the answer as text in the HTML, open every section with a standalone sentence carrying the number and repeat the same figures in a table. Those four moves cover what a page controls; the rest depends on what other websites say.

How do you optimize a website for LLMs?

Optimising a website for LLMs means optimising each page for retrieval and for lifting rather than for a model. Retrieval means crawler access, indexing and snippet eligibility. Lifting means one answer sentence per section, exact numbers in the first 30% and a table whose headers carry the judgment. LLM SEO advice beyond this, such as a separate page for every phrasing, mostly repeats what search engines already resolve.

Do I need an llms.txt file to rank in AI search?

No. Google states that its Search systems, including the generative AI features, do not use LLMS.txt or similar files, and that publishing one neither helps nor harms visibility there. Other services may read it, so treat it as optional packaging, not a ranking factor.

How long does page-level AI search work take to show up?

Crawler access and rendering fixes depend on recrawling, which Google says can take from a few days to a few weeks, while OpenAI states a robots.txt change takes about 24 hours to reach its search systems. Canlah AI re-checks answer-fit rewrites two to four weeks after release, on questions that already retrieve the site. It plans category-level visibility on a three- to six-month window because that depends on off-site pages. In Canlah AI’s client projects, new AI citations typically begin to appear about 15 days after the first batch of changes goes live.

Can Canlah AI guarantee that ChatGPT will recommend my brand?

No. Canlah AI stands behind the measurement, the sampling protocol and the work delivered, and its written bar is that AI names you at least 30% more often than at your starting baseline, re-tested every month on the same buyer questions; no ranking or #1 guarantee. It cannot control how an independent model answers a question. Treat a guaranteed AI recommendation as a reason to walk away.

Method and sources

This guide was written on September 23, 2026, and vendor documentation was read on current public pages that day without paid accounts. The citation figures come from a Canlah AI study of 60 buyer questions run against ChatGPT and Google AI Mode; it stored only the first 600 characters of each answer, so copied sentences later in an answer could not be located. The canlah.ai figures are a single-date sample from one market, not a stable rate.

• Canlah AI self-audit of canlah.ai, internal export, September 7, 2026. Crawl counts, page-level checks, crawler permissions, structured data and per-layer mention counts.
• Canlah AI cited-page format study, internal export, September 23, 2026. The 526 citations, 92 read pages, position and container findings.
• Google Search Central: Optimizing your website for generative AI features on Google Search. Grounding, query fan-out and the practices Google Search ignores.
• Google Search Central: AI features and your website. Technical requirements, snippet eligibility, preview controls and Google-Extended.
• Google Search Central: Ask Google to recrawl your URLs. The recrawl window of a few days to a few weeks.
• OpenAI crawlers. OAI-SearchBot, GPTBot and ChatGPT-User scope.
• Perplexity crawlers. PerplexityBot and Perplexity-User scope.
• Canlah AI pricing. The per-platform GEO service tiers, the free AI visibility report and the one-time AI-readability rebuild.

See where your brand stands in AI answers. Run the free AI visibility audit →

Related articles

• GEO vs SEO: What’s the Difference, and How Brand Visibility Works in the AI Era
• What Is llms.txt? A Plain Guide to the File
• How AI Engines Choose Which Sources to Cite
• Why Is My Website Not Showing in AI Overviews