How to Rank on ChatGPT, Gemini, Perplexity and Claude

There is no ranking. That is not pedantry — it is the reason most advice on this topic fails. An AI assistant does not maintain a numbered list of results for your keyword. It retrieves a set of pages, reads them, and then decides which handful to name. Two different jobs, two different failure points, and the tactics that fix one do nothing for the other.

Worse, a large share of businesses asking how to get cited have quietly told these systems not to look. The crawler that trains a model and the crawler that fetches a page to answer a live question are different bots with different names, and a robots.txt file written in 2024 to keep AI companies out very often blocks the second along with the first. That single misconfiguration removes you from the answer entirely, and no amount of content work will undo it.

This guide starts there, then works through what the controlled evidence actually supports about citability, and clearly separates it from what the market asserts without evidence.

The crawler split, from the vendors’ own documentation

Every major assistant now runs at least two distinct crawlers: one that collects content for model training, and one that indexes or fetches pages so they can be cited in a live answer. Blocking the first costs you nothing in visibility. Blocking the second removes you from the answers.

Three consequences fall straight out of that table.

First, the sensible default for most businesses is: allow every retrieval and search bot, and decide separately about the training bots. Blocking GPTBot and ClaudeBot is a legitimate commercial and rights decision that costs you no visibility. Blocking OAI-SearchBot, Claude-SearchBot or PerplexityBot is a visibility decision dressed up as a rights decision.

Second, Google-Extended is not the lever people think it is. It governs Gemini app and Vertex AI grounding, and Google states in the same documentation that it has no effect on inclusion or ranking in Google Search. Google does not name AI Overviews in that page, and AI Overviews sit inside Search, so blocking Google-Extended is not the way to opt out of them. The snippet controls covered in our guide to showing up in AI Overviews are.

Third, user-triggered fetchers behave differently from crawlers. Both OpenAI and Perplexity state, in slightly different words, that robots.txt may not govern a fetch a human explicitly asked for. If your goal is genuine exclusion rather than a gesture, robots.txt alone will not get you there.

Check your own file before doing anything else. It takes one minute, and in our experience roughly a third of the sites we audit are blocking at least one retrieval bot they meant to allow.

Retrieval, then citation: the two-stage model

When an assistant answers a question that needs live information, it decomposes the question into sub-queries, retrieves a set of candidate pages, reads them, and then names a much smaller set in the visible answer. Google describes its own version of the first half as “query fan-out” — issuing multiple related searches across subtopics rather than one search for the literal question.

The distinction matters because the two stages fail for different reasons and are fixed by different work.

You fail retrieval because you are not crawlable, not indexed in whatever index that assistant leans on, or not topically close enough to any of the fanned-out sub-queries. That is a technical and keyword-coverage problem. It is the same problem technical SEO has always solved.

You fail citation because your page was read and passed over — another source stated the same thing more directly, more specifically, or with something the model could attribute. That is a content problem, and it is the one that is genuinely new.

What the controlled evidence actually supports

There is far less rigorous research here than the volume of commentary suggests. Two studies are worth knowing properly, and they point in opposite directions on the tactics people spend most on.

The GEO paper: content changes do move citation rates

Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande published “GEO: Generative Engine Optimization” at KDD 2024. It coined the term and provided the first controlled demonstration that content edits change how often a page is cited in a generated answer. Their benchmark covered roughly 10,000 queries across nine datasets, and reported visibility gains of 22% to 41%.

The tactics that won were what the authors called epistemic authority signals: adding statistics, citing sources, and quoting credible authorities, worth up to about 40%. Tactics that read as classic SEO — keyword stuffing in particular — performed poorly.

The caveat is routinely omitted and materially changes how you should use the paper: the evaluation synthesised Google’s top five sources using GPT-3.5-turbo. That is not the 2026 production stack for any assistant. Treat the finding as strong directional evidence that specific, sourced, attributable writing is favoured over vague writing, and do not treat the percentages as a spec you can hit.

The schema study: structured data does not buy citations

Ahrefs ran the test the market has been assuming the answer to. They took 1,885 pages that added JSON-LD structured data between August 2025 and March 2026 and matched them against roughly 4,000 control pages. The result: no citation lift on AI Overviews, AI Mode or ChatGPT, and a small but statistically significant 4.6% decline on AI Overviews.

Ahrefs states its own limitation, which is worth carrying: the study looked at pages already receiving AI citations, so it measures the marginal effect of adding schema, not whether schema matters from zero. Even so, this is the only controlled test on the question, and it contradicts the most commonly sold tactic in the category. Schema markup earns its place through rich results and machine-readable clarity. It is not a citation lever.

Where the assistants genuinely differ

They are not four versions of the same system, and a single playbook will not win everywhere. The differences that actually affect what you do:

ChatGPT runs its own search index built by OAI-SearchBot, alongside undisclosed partner sources. Its search behaviour is opt-out through robots.txt, and OpenAI publishes no citation-level reporting to publishers — so there is no equivalent of Search Console for it, and any tool claiming precise ChatGPT citation data is inferring, not reading.

Google’s AI Mode and AI Overviews run on Google’s own index. Eligibility is ordinary Search eligibility, which means the work is ordinary SEO. This is the surface that matters most in Singapore, where Google holds 92.46% of search as of July 2026 per StatCounter.

Perplexity is the most transparent about its crawlers and the most explicit that PerplexityBot exists to surface and link sites rather than to train models. It is also the assistant that most consistently shows visible citations, which makes it the easiest one to audit yourself.

Claude separates training, search indexing and live user fetches into three named agents, which is unusually clear, and gives you a genuine choice: contribute to search answers without contributing to training.

Cross-platform overlap in cited domains is low — one published index put the ChatGPT and Perplexity overlap at around 11%. That figure comes from a commercial index rather than peer review, so treat the number loosely, but the direction is consistent across several analyses and it has a real implication: you cannot win all four with one page. Pick the assistant your buyers actually use and audit that one properly.

The third-party problem nobody wants to hear

Multiple 2026 analyses agree that a small set of domains — Reddit, Wikipedia, YouTube and LinkedIn among them — absorbs a large share of citations across assistants. The most-quoted specific percentages, such as Reddit accounting for around 40% of citations, come from a public-relations firm’s own index distributed by press release, not from peer-reviewed work. We use the direction and not the numbers.

The direction alone is uncomfortable and useful. It means a meaningful share of the answers about your category are being assembled from pages you do not own. Whitespark’s 2026 survey of 47 local search practitioners, published November 2025, ranked “presence of business on expert curated best-of lists” as the single highest-weighted AI visibility factor — that is practitioner opinion rather than measurement, and should be read as such, but it points the same way.

What to do with that, in decreasing order of defensibility: get accurate, current entries on the industry directories and comparison sites that genuinely rank in your category; earn coverage where a journalist or editor names you; participate honestly in communities where your category is discussed. What not to do: manufacture Reddit threads about yourself, or buy placement on a list that exists only to sell placement. Both are visible, and both are the kind of thing Singapore’s advertising and consumer-protection rules take a dim view of when the commercial relationship is undisclosed. Earning links honestly and earning citations honestly turn out to be nearly the same discipline.

A practical sequence

If you do these in order, you will have done the defensible version of this work.

One. Read your robots.txt and confirm every retrieval bot in the table above is allowed. Decide training bots separately, on rights grounds.

Two. Confirm your key pages are indexed and snippet-eligible. An excluded page cannot be retrieved by anything.

Three. Write down the ten questions your buyers actually ask, in their words. Ask each assistant those ten questions and record what comes back: are you named, is a competitor named, is the description of your category accurate. This baseline is worth more than any tool subscription.

Four. Fix the answers that are wrong about you. An assistant describing your service incorrectly is usually reading an outdated third-party page, and correcting that page is faster than out-publishing it.

Five. Restructure your best pages into self-contained, question-headed sections with the answer first. This serves fan-out on every platform at once.

Six. Add what a model cannot generate: your own numbers, your own methodology, named people with real credentials, dated observations from your own work. This is the tactic with the strongest experimental support behind it.

Seven. Re-run the ten questions monthly and log the change. Nothing here is measurable in a week.

Measuring it, and the trap in GA4

GA4 now carries a native AI Assistants default channel covering traffic from sources like ChatGPT, Gemini, Deepseek, Copilot and Grok. It requires the medium to match ai-assistant exactly and the referrer to be on Google’s maintained list.

Two limits to hold in mind. Traffic from Google’s AI Overviews and AI Mode is explicitly excluded from that channel and classified as Organic Search, so the channel will never show you AI Overview performance. And a large share of genuine assistant referrals arrive with no referrer at all and land in Direct, which means the channel systematically under-counts. Our guide to the GA4 reports worth checking covers where it lives, and the discipline of not celebrating the wrong number is in our piece on vanity metrics.

Given all that, the manual log of ten questions remains the most reliable instrument available. It is unglamorous and it works.

The Singapore angle

Singapore is an unusually good market for this work and an unusually bad one for over-investing in it. Stanford HAI’s 2026 AI Index puts local consumer generative-AI adoption near the top globally at roughly 61%, against about 53% worldwide, so your buyers genuinely are asking assistants. But Google still handles 92.46% of Singapore search, so the assistant that matters most here is the one built into the results page rather than a standalone chatbot.

The proportionate response for a Singapore SME: allow the retrieval crawlers, run the ten-question audit quarterly, and fund the content work inside your existing SEO programme rather than as a separate retainer. If a proposal prices chatbot visibility as its own product, ask which activities in it are not already SEO. The list is usually short. Our wider view is in the Singapore AI SEO guide, and the terminology minefield is unpicked in GEO vs AEO vs LLM SEO.

Frequently asked questions

Can I pay to be recommended by ChatGPT?

Not for the citations in an organic answer. Those are selected by the system from retrieved pages. OpenAI has been testing advertising formats inside ChatGPT, but those are labelled placements and are a separate purchase from being named as a source. Anyone offering guaranteed ChatGPT recommendations is selling something they do not control.

If I block GPTBot, will I disappear from ChatGPT answers?

No. GPTBot collects content for training foundation models. The bot that governs whether you appear in ChatGPT’s search answers is OAI-SearchBot, and OpenAI states that sites blocking it “will not be shown in ChatGPT search answers, though can still appear as navigational links”. You can decline to be training data and still be citable — but only if you keep those two rules separate in your robots.txt.

Does adding FAQ schema get me cited more often?

The only controlled test says no. Ahrefs matched 1,885 pages that added JSON-LD against about 4,000 controls and found no citation lift, with AI Overview citations running roughly 4.6% below control. Writing genuine question-and-answer sections does help, because assistants match sub-questions to passages. The markup on top of it is a separate claim with weak support.

How long does it take to start appearing in AI answers?

Retrieval fixes such as unblocking a crawler can take effect within days — OpenAI notes robots.txt changes take about 24 hours to process. Content and third-party changes behave like SEO: weeks to months, depending on how often those sources are recrawled and how competitive the question is. Expect no meaningful signal inside a month.

Do I need a separate agency for AI visibility?

Almost certainly not. Google’s own documentation states there are no additional requirements or special optimisations for its AI features, and the tactics with real evidence behind them — crawler configuration, page structure, original sourced content — sit inside any competent SEO scope. What is reasonable is to add monitoring and a crawler policy to an existing programme. Our guide to choosing an SEO agency in Singapore covers the questions worth asking.

Should I create an llms.txt file to help assistants find my content?

Not for citation purposes. No major provider has committed to reading llms.txt in production, and Google’s documentation says you do not need AI text files. There is one genuine use case: coding and agent tooling does consume it, so if developers read your documentation there is an argument for publishing one. That argument has nothing to do with being named in a search answer.

Where this leaves you

Strip out the marketing and the work reduces to three things. Let the retrieval crawlers in, because that is binary and free. Structure your pages so that a single sub-question can be answered from a single section, because that is how the systems read. And publish material containing something only you can supply — your own figures, your own method, your own named people — because the one controlled study on content effects says that is what gets attributed.

Everything beyond those three is either unproven or is ordinary SEO with a new label on the invoice. That is not a disappointing conclusion; it means the investment you have already made is doing more work than the market is telling you.

If you want the crawler audit, the ten-question baseline and the content work run properly, our AI SEO service in Singapore does it as part of an SEO engagement. Related guides: how to show up in AI Overviews for Google’s surfaces, AI SEO tools in 2026 for what monitoring costs, and AI SEO tools for small businesses if you are running this with a small team. Our results for other Singapore businesses are in the case studies.

Last updated 4 August 2026. Written by Adrian Tan and the SDM team. Sources: OpenAI crawler documentation (developers.openai.com); Anthropic crawler documentation (support.claude.com); Perplexity crawler documentation (docs.perplexity.ai); Google Search Central documentation on AI features and on Google-Extended; Aggarwal et al., “GEO: Generative Engine Optimization”, KDD 2024; Ahrefs controlled study of schema markup and AI citations (2026); Whitespark 2026 Local Search Ranking Factors survey (November 2025); Google Analytics 4 default channel group documentation; StatCounter Singapore search engine market share, July 2026; Stanford HAI 2026 AI Index.

Related guides

The idea underneath all of this has a name and a paper behind it: what generative engine optimization actually is covers the KDD 2024 research, what it measured, and which parts of the commercial version are not supported by it. For the full cross-surface programme, see SEO for AI search. If you are weighing up whether to pay for monitoring, ChatGPT SEO tools separates the visibility trackers from ChatGPT-as-a-tool and sets out a free sixty-minute check to run first.