GEO Playbook: How to Get Cited by ChatGPT, Claude, Gemini and Perplexity in 2026
Sandy
Head of Content · Block AI
Quick Answer
Generative Engine Optimization (GEO) in plain English: how AI assistants pick their sources, what to do on your own site, where to get external mentions, and how to measure whether it is working. Based on research published through October 2026.
The core finding
When an AI assistant searches the web, it does not search on the user's exact question. It rewrites the question into several short sub-queries of its own and then names the brands that appear consistently across multiple results.
A citation in a source list rarely drives a click: users click source links in Google's AI Overviews only about 1% of the time (Pew Research, 2025). Microsoft Copilot cited one site 17,400 times in three months and delivered 167 sessions from Bing in the same period (GEOforge, Aug 2026).
The goal is not citations. The goal is your brand name appearing in the answer text so the reader forms an intent and comes to you directly. Measure it by asking new customers how they heard about you, not by counting source links.
The order of work: open your site to AI crawler bots, build pages worth linking to, earn mentions on the platforms each assistant already trusts, then measure through quarterly manual checks and customer surveys.
How each assistant builds its answer
Sub-queries
ChatGPT breaks one user question into 1 to 7 search queries of 5 to 7 words each. Google's AI Mode uses 10 to 40 sub-queries per answer (Mike King, iPullRank). Claude adds the current year to 94% of its sub-queries and in about one in six queries inserts a category leader's name unprompted (Searchable, Aug 2026).
What the assistant sees
At the search stage, ChatGPT sees the URL, title, description, and a text snippet. Claude's search tool reads only the title and URL, then opens the page, strips it to plain text, takes the first 100 KB, and summarizes it with a smaller model. The extracted quote is capped at roughly 125 to 150 characters (Anthropic documentation breakdown, Reddit). AI crawlers other than Google's do not execute JavaScript, so they cannot see text that a script renders after page load (Vercel / MERJ, 2025).
How many sources it uses
Per answer: ChatGPT averages 4 source links, Claude 3.5, Perplexity around 13, Grok 0.7 (Max Planck Institute, Sep 2026). Assistants rarely cite more than 2 pages from the same domain.
It answers from cached copies
ChatGPT builds its own index. Bing Webmaster Tools for one site showed 57,000 citations pointing to just 19 pages: the assistant retrieves a saved copy, it does not open the page fresh each time. This is why freshness and recrawl notifications matter.
Which sources each assistant prefers
| Assistant | Primary index | What it cites most | Key driver |
|---|---|---|---|
| ChatGPT | Own index, Bing, Google | Brand homepages 44.7% of citations; "best of" listicles 43.8% of cited pages. Reddit nearly vanished after August 2026. | Bing and Google positions, brand mentions in press, homepage quality |
| Gemini / AI Overviews / AI Mode | Google index | AI Overviews: brand sites 59.8%. AI Mode: half its citations come from just 71 domains; YouTube is the top external source. | Standard Google SEO, industry publications, YouTube |
| Perplexity | Own crawler and real-time search | Reddit 21.6%, YouTube 20.8% within top 50 domains; brand sites 28.9%. More likely to surface new brands than ChatGPT or Gemini. | Videos, Reddit, review sites, comparison articles |
| Claude | Brave Search | 79.2% of citations from Brave's top 10; 34% from Google's top 10. Prefers editorial long-form, documentation, and lists. Rarely cites Reddit. | External coverage, Brave ranking, pages open to Googlebot (Brave follows Google's crawl path) |
| Grok | Own search and X posts | Fewest citations per answer (0.7). Primarily cites Reddit as external source. An X post with a list was cited within 4 days. | Only audit if your customers actively use X |
Only 23 domains out of 253 frequently cited domains are cited by all five systems simultaneously (Vercite, Jun 2026). Target the assistants your customers actually use.
Timeline: what to expect and when
These are individual observed cases, not averages.
| What | Observed timeframe |
|---|---|
| New content from an established site appearing in ChatGPT, Perplexity, Claude | ~3 hours |
| Short article earning first citations | ~21 days on average |
| How-to guide or tutorial earning first citations | ~4 months on average |
| New brand appearing in Perplexity after external articles published | Days to 2 weeks |
| Deleted page disappearing from Grok and Perplexity | Almost immediately |
| Deleted page disappearing from ChatGPT and Claude | 3 to 4 weeks |
| Full page rewrite reclaiming first position | ~2 weeks |
Half of the pages that assistants cite were published within the past 13 weeks (Profound, Sep 2026). A citation's visibility drops by half roughly every 4.5 weeks on average (Scrunch + Stacker, 3.5M citation events). Update key pages at least once a month.
Week 1: technical setup
Before worrying about content, make sure the assistant's crawlers can actually reach your site.
1. Check robots.txt
Open your robots.txt and verify these user agents are allowed: OAI-SearchBot and ChatGPT-User (OpenAI), Claude-SearchBot and Claude-User (Anthropic), PerplexityBot and Perplexity-User, bingbot and Googlebot. Also check for a "# BEGIN Cloudflare Managed content" block: Cloudflare can silently append denials for GPTBot, ClaudeBot, and Google-Extended. If present, disable managed robots.txt in Cloudflare settings.
Do not add a rule that allows only named bots. Brave Search does not announce itself with a distinct crawler name, and that rule would block Brave entirely, cutting off Claude in the process.
2. Verify the crawler actually gets a page
From any cloud server (not your office IP), request 5 key pages with each bot's user-agent string using curl. The response should contain actual text, not an empty template, a captcha, or a 403 error. A 403 on a cloud IP does not necessarily mean the real crawler is blocked: confirm with server logs or Bing Webmaster Tools before changing anything.
3. Add a date string and test
Add a line with today's date to a key page. Ask ChatGPT, Claude, Gemini, and Perplexity to open that page and read back the date. If they cannot, check your server logs to see whether the crawler even arrived, and whether your server responded correctly.
4. Set up measurement infrastructure
- Submit your sitemap to Bing Webmaster Tools and enable IndexNow (supported by Bing and other engines, not Google).
- In GA4, create a custom channel group called "AI Assistants" using sources chatgpt.com, perplexity.ai, gemini.google.com, claude.ai, copilot.microsoft.com, and grok.com, placed above the Referral channel. Some ChatGPT links arrive without utm_medium, so match on source only.
- Add "How did you hear about us?" to your registration or checkout form with explicit options for each AI assistant alongside non-AI options. Without non-AI choices the AI share looks inflated.
- Build a list of 20 customer questions for monthly manual checks. At least 18 of the 20 should not include your brand name.
Weeks 2 to 4: your own site
Homepage
Between a quarter and 60% of ChatGPT referral sessions land on the homepage (Similarweb, Aug 2026). The first screen without scrolling should answer: what this is, what it costs, how to pay, and how to start. Do not turn the homepage into a FAQ page; build separate pages for questions.
Prices as HTML text
Pages with prices are cited more often than other formats (Discovered Labs, 2M citations). A PDF price list, a pricing widget loaded by JavaScript, and prices inside images are all invisible to AI crawlers. Verify by pressing Ctrl+U and searching for the price in the page source.
Answer in the first words
Google's AI Overviews pull 11% of their cited snippets from the first 15 words of a page; almost nothing is pulled from past 500 words (Surfer, Jun 2026). Each key page should open with a 40 to 60 word self-contained answer, then expand with sub-headings.
One page per question
Give each major customer question its own page. When the page title matches the question at least halfway, ChatGPT cites it in 20% of relevant answers; when it barely matches, that drops to 9% (AirOps, DesignRush). Use a readable URL with 5 to 7 ordinary words, not IDs or codes. Write the meta description as a direct summary of the answer.
Sentences that survive extraction
Claude's cited snippet is about 150 characters. Each key claim must be complete in one sentence: subject, number, and condition. Phrases like "as mentioned above" do not survive when the sentence is extracted.
Your own data
Ramp turned spending data from 40,000 customers into monthly-updated charts with an economist's commentary. Sessions grew by nearly 150% and Ramp started appearing alongside major publications in AI answers (Ramp, Nov 2025). If your business generates proprietary metrics, publish a page for them and update it on a fixed schedule.
Honest comparisons
A small SaaS called FluoTest wrote a direct comparison presenting itself as a free alternative to ScoreApp. Over three months it earned 131 users with no advertising; roughly half the sign-ups came from ChatGPT (founder account, Reddit). One or two honest comparisons with a named competitor work. Hundreds of templated versions do not: Google has begun downranking mass-produced self-promotional lists since early 2026.
Months 2 to 3: external pages
Brand mentions across the web correlate with AI visibility at 0.66 to 0.74 (Ahrefs, 75k brands, Dec 2025), stronger than backlink counts. A single article on a domain the assistant already trusts regularly outperforms 20 posts on unfamiliar domains (Fudan University, NDSS 2027).
The target for one quarter: 15 to 20 external pages on domains the assistant actually cites, each mentioning your brand by name alongside your price and what makes you different.
How to choose platforms
Run your 20 customer questions in each assistant you care about. Record which domains appear as sources. Work only those domains. A platform that never appears in those answers is not worth your time for GEO purposes, whatever its general authority.
"Best of" lists
Authors of existing round-up lists are your highest-leverage target. Prepare a pre-written paragraph about your product and a screenshot, then reach out. These lists account for 43.8% of pages ChatGPT cites (Tim Soulo, synthesis of 14 studies) and 36.4% of Claude's citations. Being second on many lists beats being first on one (Leo Poitvein).
Review platforms
For B2B software, five platforms (Gartner Peer Insights, G2, Capterra, Software Advice, TrustRadius) account for 88% of review-site citations in Google's AI Overviews (SE Ranking). A bad rating on Trustpilot has outweighed ten paid reviews: one business saw ChatGPT actively warn users against buying after a poor Trustpilot score drove the recommendation (Leo Poitvein). Respond to every negative review and ask satisfied customers to write externally.
YouTube
Works for Gemini, Google's AI Mode, and Perplexity. One tested format across 12 videos showed Gemini citing 4 of the 12 after 6 weeks: the cited videos answered the question within the first 60 seconds. Generic slide-based AI-voiced videos earned zero citations across 493 videos tested by one agency (PushLeads, Aug 2026).
Format for maximum citability: one question per video, answer in the first minute, live product demo rather than slides, title that includes your brand and a price or key number, an SRT subtitle file, chapters with timestamps, description with prices and a UTM-tagged link.
Currently cited heavily by Perplexity and Google's AI Overviews; nearly absent from ChatGPT after August 2026. Works best for local services in markets without established authority sources. Authentic long-term participation works; placing promotional mentions gets accounts banned. One experiment showed that multiplying account karma by 23x added zero extra citations (Reddit, Jun 2026).
Press mentions
ChatGPT explained in one case study that it did not recommend a brand because it could not find the brand on trusted sources including press, review sites, and industry directories (Head Promo, Aug 2026). Two-dollar and six-dollar press releases about a fictional event landed in Google's AI Overview as though the event had happened. That same mechanism works in your favor for factual announcements on legitimate newswires, but it also means competitors can introduce wrong information about you; publish your facts on your own site and on external pages using the same wording.
What does not work
- llms.txt: not used by any major AI assistant. Google explicitly ignores it. In a 3-month controlled experiment, no AI crawler requested the file or the pages listed in it (Reboot Online, Mar 2026).
- Schema markup for citation gains: a controlled study on 1,885 pages found no increase in AI Overview, AI Mode, or ChatGPT citations (Ahrefs, May 2026). Schema is still useful for correcting factual errors about your brand in AI answers, but it is not a citation lever.
- Markdown copies of pages: in a Profound controlled study, Markdown versions of pages showed no meaningful difference in bot visits compared to HTML.
- Buying backlinks or using PBN networks: Google neutralizes them, and external mentions correlate with AI visibility far more than link counts.
- Mass-produced "best of" lists with your brand first: Google has been downranking these since early 2026; ClickUp lost significant traffic after publishing thousands of them.
- Reddit karma manipulation: no citation lift observed in any published test.
How to measure
Measure by conversions, not citations.
Monthly manual check (minimum): run your 20 questions in ChatGPT, Gemini, Perplexity, and Claude. Ask each question 3 to 5 times in a fresh private session. Record: was the site cited, was the brand named, which competitor domains appeared, what sub-queries did the assistant reveal. That is 240 to 400 answer records per month across four assistants. ChatGPT's sources change faster than others: check it every 2 weeks if you can.
Quarterly metrics to compare against your baseline:
- Conversions attributed to AI channels in GA4
- Share of "how did you hear about us" survey responses naming an AI assistant
- Percentage of manual-check answers that name your brand on questions without your brand name
- Number of external pages on cited domains where your brand appears with a price and a differentiator
Count a metric as improved only if it holds across at least 2 consecutive checks. The primary signal is customers: while they are scarce, the early leading indicator is the brand-mention rate in manual checks.
Bing Webmaster Tools AI Performance report: shows Copilot and Bing AI Overview citations; a useful indirect proxy for ChatGPT since Bing is one of ChatGPT's primary sources.
Server logs: identify visits from OAI-SearchBot, ChatGPT-User, Claude-SearchBot, Claude-User, PerplexityBot, Perplexity-User. Cross-reference IPs against the published IP lists from OpenAI, Anthropic, and Perplexity: in one 2-month log analysis, 76% of visits claiming to be GPTBot and 39% claiming to be PerplexityBot came from IPs not on the official lists. Fake crawlers exist and search for password files.
The one-paragraph summary
AI assistants name brands whose facts appear consistently across multiple trusted sources. Open your site to their crawlers (check robots.txt and Cloudflare). Publish prices as HTML text with a direct answer in the first 50 words of each page. Build one page per major customer question and update key pages every month. Then earn mentions on the specific domains each target assistant cites by reaching out to list authors, building a review presence, and creating YouTube content for Gemini and Perplexity. Measure by asking new customers how they found you and by running your question list manually each month. Assess quarterly.