How to Rank in ChatGPT, Gemini, Perplexity and Claude: The 2026 GEO Playbook

There is no ranking in ChatGPT. There is retrieval, synthesis, and citation. A model reads a handful of sources it can reach, writes an answer, and names some of them. Your job is to be reachable, quotable, and corroborated, in that order.

We build for this on client sites and on our own products. Our brand-audit tool received its first ChatGPT citation 48 hours after launch with no advertising, and the reasons were structural rather than clever. This is the working playbook as of mid-2026, including the parts that stopped working.

If you have not read SEO vs GEO, start there for the definitions. This article assumes them.

How an AI answer is actually assembled

Four steps, roughly the same across engines.

1. Fan-out. Your one question becomes several internal searches. Google describes this behaviour openly in its generative AI optimization guide. Ask "best CRM for a small Indian agency" and the system quietly runs half a dozen narrower queries.
2. Retrieval. Each sub-query pulls candidate pages from an index. Which index depends on the engine.
3. Synthesis. The model reads what it retrieved, mostly the passages that look like answers, and composes a response.
4. Citation. It attaches sources to claims, favouring pages whose text supports the sentence it just wrote.

Every optimization below maps to one of those four steps. If you cannot say which step a tactic affects, it is probably decoration.

Which index each engine leans on matters more than any content trick:

• ChatGPT uses Bing's index for live retrieval, plus its own crawl. Bing Webmaster Tools verification is therefore mandatory, and it is the single most skipped ten minutes in Indian marketing.
• Google AI Overviews and AI Mode use Google's index. Classic Google SEO is the entry ticket.
• Perplexity runs live retrieval across its own crawl and third-party indexes.
• Claude uses its own retrieval with web search.
• Copilot rides Bing, same as ChatGPT's live path.

Two indexes, Google and Bing, cover the large majority of what gets cited. Most businesses are in one and absent from the other.

Step 1: Be fetchable, and prove it

Nothing else counts if the crawler gets a 403.

Run this against your own site:

curl -A "GPTBot/1.0" https://yoursite.com/
curl -A "OAI-SearchBot/1.0" https://yoursite.com/
curl -A "ClaudeBot/1.0" https://yoursite.com/
curl -A "PerplexityBot/1.0" https://yoursite.com/

You are looking for your real HTML. A 403, a 503, a captcha page, or an empty shell means your CDN, WAF, or bot-protection layer is blocking the exact agents you want. This is by far the most common cause of a site having zero AI visibility, and it is usually a setting nobody chose deliberately.

Then check that content exists without JavaScript. If the page body only appears after a client-side fetch, assume a meaningful share of crawlers see an empty page. Server-render anything you want quoted.

Finally, decide crawler policy on purpose. Allowing GPTBot while blocking ClaudeBot is a decision; inheriting whatever your hosting default is, is not.

Step 2: Be in both indexes

Verify Google Search Console. Then verify Bing Webmaster Tools, submit the sitemap, and use its URL submission. Bing indexing is the gate to ChatGPT and Copilot retrieval, and Bing's share of classic search is irrelevant to that fact.

Wire IndexNow so new and updated URLs are pushed the moment they publish rather than waiting to be discovered. It is a small amount of code, it works with Bing, Yandex, Naver, and Seznam, and it compresses discovery from days to hours. Our own publishing pipeline pings it on every Sanity webhook, which is why new articles turn up in Bing the same afternoon.

Step 3: Write in passages, not essays

Retrieval works on chunks. The rewrite rule that matters:

• Question-shaped H2s and H3s. Use the phrasing a person types, not the phrasing a brand prefers. "What does a brand identity cost in India" beats "Investment".
• Answer in the first sentence under each heading. No throat-clearing, no context-setting paragraph before the answer.
• Keep each answer self-contained, roughly 60 to 120 words, understandable if lifted out of the page with no surrounding text. That is exactly what happens to it.
• One idea per paragraph, three sentences or fewer.
• Repeat the entity name rather than leaning on "we" and "our platform" throughout. A lifted paragraph that says "we" tells the model nothing about who to credit.

This is the same discipline that makes an FAQ work, applied to the whole document. A long article structured as fifteen answers gets fifteen chances at a citation; the same article structured as one argument gets one.

Step 4: Put the facts in the text

Models quote text. They do not read your infographic, your PDF price list, your carousel, or the number rendered inside a JavaScript-driven accordion that never opens for a bot.

If it matters, write it as a sentence: what it costs, how long it takes, where you operate, what is included, what is not, when it was last verified. Indian businesses in particular hide pricing as a matter of habit, then wonder why AI answers about their category name three competitors who publish ranges.

Step 5: Statistics, quotes, and named sources

This is the one piece of GEO advice with a controlled experiment behind it. The Princeton GEO study, presented at ACM SIGKDD in 2024 and run across roughly 10,000 queries, tested nine content modifications. The three that produced the largest visibility gains, in the range of 22% to 41%, were adding verifiable statistics, adding credible quotations, and citing reliable sources.

Keyword stuffing, by contrast, produced nothing.

Practical target for a 2,000-word page: ten to twenty verifiable references, five to ten specific statistics with named sources, and at least one quoted expert with full attribution. Not decoration. The claim and its source travel together into the answer, which is what makes the model comfortable repeating it.

Step 6: Treat freshness as maintenance, not publishing

Cited pages in 2026 skew measurably newer than the organic top ten, roughly a quarter fresher in citation-factor analyses, and individual pages lose visibility within weeks of going stale.

The operational version of that fact:

• Put a visible last-updated date on substantial pages, and make it true.
• Refresh your top twenty commercial pages every 30 to 90 days: update numbers, replace dead sources, add whatever changed.
• Re-ping IndexNow on update, not only on publish.

Half the AI visibility gains we see on client sites come from maintaining twenty pages properly rather than publishing forty new ones badly.

Step 7: Fix the entity graph off your own site

This is the step most teams skip and the one that decides the hardest queries. Models cite what they are confident about, and confidence comes from several independent sources agreeing. Analyses through 2026 put roughly 84% of AI citations on earned media rather than the brand's own domain, and find brand mentions correlating with AI visibility around three times more strongly than backlinks.

Get the same facts, written the same way, into at least eight of these:

• LinkedIn company page and founder profiles
• Crunchbase
• Google Business Profile and Bing Places, both, with matching NAP details
• YouTube, which shows up repeatedly as a top correlate of AI visibility
• Wikidata, where eligible
• Industry directories and association listings that are real
• Trade press, podcasts, and newsletters in your category
• Your own about page, with the entity written in one clean paragraph

Same founding year, same legal entity, same one-line description, same location. Where models find contradictions they hedge, and hedging reads to a buyer as "this company might not exist".

Step 8: Earn mentions where models look

Beyond profiles, citations come from places models trust for opinion: community threads, comparison content, listicles, review sites, and trade coverage. Being in the "best X in India" articles that already get cited is often faster than trying to outrank them.

Do this the honest way. Contribute where you genuinely have expertise, under real identities, and pitch real stories to real publications. The shortcut version, burner accounts seeding threads, is now actively hunted: Reddit removed roughly 25,000 spammy posts and comments a day in Q1 2026 and deployed LLM-based detection for coordinated behaviour in July 2026. When the post gets deleted, the citation dies with it, and you have bought a liability.

Step 9: Schema, scoped honestly

Ship structured data: Organization, WebSite, Service or Product, BreadcrumbList, FAQPage, Article where it applies. Keep it consistent with the visible page.

What it does: helps machines parse entities and relationships, earns rich results in classic search, and those results feed the retrieval layer.

What it does not do: force a model to quote you. Google's 2026 guidance is explicit that no AI-specific schema is required for its generative features. Schema is hygiene with real second-order value, not a citation lever. Anyone pricing it as the core of a GEO engagement is overcharging for an afternoon of work.

Step 10: Own the query shapes AI answers get built from

Generative answers are disproportionately built from a few page types, because those are what the fan-out queries look for:

• Comparison pages. "X vs Y", including comparisons where you lose on some dimensions. Balanced comparisons get cited more than one-sided ones.
• Pricing and cost pages with real ranges and what moves them.
• Alternatives pages. "Alternatives to [category leader]", written fairly.
• Definitional explainers for the terms your buyers are confused by.
• Process pages. What actually happens, week by week, with timelines.

If you publish nothing in these shapes, you are absent from the answer regardless of how good your homepage is.

How to measure it

There is no rank tracker, because there is no rank. Build a prompt panel instead.

1. Write 30 to 50 prompts a real buyer would type, spread across research, comparison, and purchase intent.
2. Run them monthly across ChatGPT, Google AI Mode, Perplexity, and Claude. Same prompts, same order, logged.
3. Record three things per prompt: were you named, were you cited with a link, and was what the model said about you accurate.
4. Track the trend, not the snapshot. Answers vary run to run, so a single result proves nothing and a shift across forty prompts proves something.

Alongside that, segment AI referrals in analytics by source, and read your server logs for GPTBot, OAI-SearchBot, ClaudeBot, and PerplexityBot hits. Crawler visits are the leading indicator; citations follow, and referrals follow those.

Paid visibility tools automate the prompt panel. They are reasonable instrumentation, and none of them do the work. Buy one after the work is underway, not instead of it.

The full method, including how to build the panel, the four numbers worth reporting, and how to avoid optimizing the measurement instead of the market, is in how to measure AI visibility.

The 30-60-90

Days 1 to 30, plumbing and baseline. Curl tests for every crawler. Fix whatever blocks them. Verify Google Search Console and Bing Webmaster Tools, submit sitemaps, wire IndexNow. Build the prompt panel and record the baseline, including what the models currently get wrong about you. Clean the entity across your top eight external profiles.

Days 31 to 60, structure. Rewrite the top ten commercial pages into passages: question headings, answer-first, self-contained, real numbers, named sources. Add schema. Publish the two or three missing page shapes, usually pricing, comparison, and process.

Days 61 to 90, authority and rhythm. Start earning mentions: trade press, podcasts, communities, video. Set the refresh cadence on the top twenty pages. Re-run the prompt panel and compare against baseline. By now the plumbing fixes have shown up, the structural work is landing, and the entity work is beginning to.

Anyone promising results faster than that is either fixing a blocked crawler, which is genuinely fast, or lying.

What we did on our own product

Brandauditor.ai launched with the plumbing already correct: explicit crawler allows tested by curl, Bing verified before launch, IndexNow wired to the deploy hook, entity written identically across external profiles, pages structured as answers, category words in the product name itself. First ChatGPT citation arrived 48 hours after public launch with zero paid advertising, and LLM citations became its primary acquisition channel.

None of that was a growth hack. It was ten unglamorous decisions made before launch instead of after. The full breakdown is in how Brandauditor hit its first ChatGPT citation in two days and the product-side version in building products designed to be cited by LLMs.

Before you spend on any of this, read the mistakes article. Several tactics that were standard advice in 2025 are now either useless or harmful.

FAQ

How long does it take for ChatGPT to cite my site?

If the blocker was crawler access or missing Bing indexing, days. For a normally functioning site, structural content changes typically surface in AI answers within two to six weeks after recrawl. Competitive queries that depend on third-party reputation take three to six months, because they wait on sources you do not control.

Do I need an llms.txt file?

No. Google's May 2026 guidance states plainly that llms.txt files, special machine-readable formats, and AI-specific rewrites are not needed to appear in its generative features, and no major AI provider has committed to using the file in production retrieval. We still ship one because it costs an afternoon and does no harm, but it is not a strategy and should never be a line item on an invoice. The full reasoning, including how to write one properly if you do ship it, is in do you need an llms.txt file.

Does adding schema make ChatGPT quote me?

Not directly. Schema helps machines parse your entities and earns rich results in classic search, which feeds the indexes generative engines retrieve from. That is real but indirect. Treat structured data as hygiene, and treat any proposal that centres on schema as the main deliverable with suspicion.

How do I check whether AI crawlers can actually see my site?

Run curl with each crawler's user agent, GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, and confirm you get real HTML back rather than a 403, a captcha, or an empty shell. Then check your server logs for those agents over the last 30 days. No crawler hits means no citations, and no amount of content work changes that until the access problem is fixed.

Do I need a blog to get cited?

You need pages that answer questions, which usually means articles, but volume is not the lever. Twenty well-maintained pages covering pricing, comparisons, process, and the real questions of your category outperform two hundred thin posts. Publishing at scale with generated content is actively counterproductive, since commodity content is the category models are least likely to cite.

Can I pay for placement in AI answers?

Not in citations, as of mid-2026. There is no paid slot in an AI Overview citation or a ChatGPT source list, and any vendor offering guaranteed placement is either misdescribing advertising products or selling something that does not exist. Sponsored placements inside AI products are being tested by the platforms themselves; that is advertising, and it is labelled.

How often should I refresh a page?

Top commercial pages every 30 to 90 days, everything else as facts change. Update the numbers, replace dead links, add what is new, and set a visible last-updated date that is honest. Cited pages run measurably fresher than organic results, so maintenance is a visibility lever rather than housekeeping.

We ship this stack as part of website engagements and document the methodology publicly on our citations page. To scope a project, start here.

NOW Media is a Bangalore creative studio founded in 2019, a brand of Bleep Design Private Limited. We run this playbook on our own products before selling it to anyone.