---
title: "ChatGPT Search Optimisation"
description: "How ChatGPT Search retrieves and cites pages, what OpenAI documents and does not, and what the stored 200-character snippet means for your H1."
tldr: "ChatGPT Search, the web search inside ChatGPT that cites its sources, retrieves through several pipelines: 2026 traffic studies found OpenAI's own index the default, with Microsoft's Web IQ serving 'some' answers. OpenAI's only published guidance is that any public site can appear if OAI-SearchBot is not blocked; it names no citation signals. The one measured mechanism is the stored snippet: Resoneo found it is cut just past 200 characters, anchored on the H1 in 83.6% of cited pages. Citation behaviour shifts between GPT-5 minor versions: the GPT-5.6 rollout of August 2026 roughly tripled searches per prompt and doubled sources per answer, so a breadth figure belongs to its model version."
publishDate: 2026-04-25
updatedDate: 2026-09-28
author: "Liam Hayward"
---
ChatGPT Search is OpenAI's web-connected search interface, launched in late 2024 and progressively expanded throughout 2025. Across 1,215 of Conductor's US enterprise customer domains in 10 industries, between May and September 2025, ChatGPT accounted for 87.4% of all AI referral traffic, making it the dominant AI search channel by a considerable margin.[^6] It allows ChatGPT to retrieve current information from the web and cite sources within its responses. From an SEO perspective, it is a separate retrieval surface from Google AI Overviews, with overlapping but distinct mechanics. This page is about making your site visible when ChatGPT answers a query; it is a separate question from whether [ChatGPT can do SEO](/ai-search/chatgpt-for-seo/) as a workflow tool.

## How has ChatGPT's citation behaviour changed by model version?

Citation behaviour moves between GPT-5 minor versions, in both directions, so a breadth or share figure belongs to the model version that produced it.

The current state is GPT-5.6, rolled out in early August 2026. Nectiv re-ran roughly 4,000 prompts against a 2025 baseline and found searches per prompt up from 2.17 to 7.61, a `site:` operator in 64% of them, and `official` and `gov` among the five terms ChatGPT most often adds to a search the user never typed, displacing the `reviews`, `free` and `comparison` of a year earlier.[^13] Sources cited per answer roughly doubled over the same rollout, which is why listicles and comparison pages lost around half their citation share without losing citations in absolute terms; [the fan-out shift](/seo-news-updates/chatgpt-fan-out-query-shift-listicles/) sets out that arithmetic.

The versions before it did not move in one direction. GPT-5.3 Instant cut unique domains per answer from roughly 19 to 15.[^3] GPT-5.4 introduced heavy `site:` use and cited brand-owned sites in 56.8% of responses; GPT-5.5 Instant, the default from May 2026,[^7] pulled both back, to 12.6% and 47.2%.[^14] The middle two rest on 50 prompts from one account on one day, and each panel measured against its own baseline, so read every figure here as the direction of a change under that version, not as a rate that carries across to the next.

A surface you do not own can lose most of its citation share the same way. Promptwatch measured Reddit's share of ChatGPT Search citations falling from 3.8% to 0.5% in the week after 8 August 2026, with the size of the drop provisional;[^15] [Reddit's citation collapse](/seo-news-updates/chatgpt-reddit-citations-drop-august-2026/) has what the other panels recorded. That is the argument for treating a third-party profile as worth having for the audience it earns you directly, with any ChatGPT citation counted as upside.

## How does ChatGPT Search retrieve content?

ChatGPT Search draws on several retrieval sources and applies OpenAI's own ranking and citation logic. OpenAI runs its own search crawler, and says OAI-SearchBot "is used to surface websites in search results in ChatGPT's search features".[^17] When a user asks a question that benefits from current information, ChatGPT issues one or more search queries, retrieves results, reads relevant pages, and synthesises a response.

OpenAI does not document which providers sit behind those results, and the 2026 evidence does not support the long-standing assumption that Bing is the main one. Independent inspection of ChatGPT's network traffic found web results arriving through several distinct pipelines. In Chris Green's sample of 1,000 prompts, published in July 2026 before the GPT-5.6 rollout, one pipeline labelled "labrador" was the primary search source 88.1% of the time, with the others corresponding to the data providers Bright Data and Oxylabs; that sample was itself collected through Bright Data's scraping service, which he flags as a limit on how far it generalises.[^8] Resoneo identifies labrador as OpenAI's in-house index, not rebadged Bing, and the Bright Data pipeline as scraped Google results, used mostly for news and current-events questions; a separate Bing pipeline was reported on a Team account and never appeared on its own accounts.[^11] The Bing assumption dates from tests like Seer Interactive's in early 2025, which found "87%+" of about 500 citations matched Bing's top organic results,[^2] an overlap rather than a supply route. Treat Bing indexing as one route among several, and OAI-SearchBot access as the one OpenAI documents.

Those pipeline findings are dated. OpenAI removed the response-stream field that named the retrieving pipeline around 21 July 2026, so the split between pipelines can no longer be read directly from the traffic.[^11] The distributions reported in the mid-2026 teardowns describe how retrieval was routed that month, not a stable architecture.

In June 2026 Microsoft launched Web IQ, an AI-native grounding API built on Bing's index.[^5] Jordi Ribas, Microsoft's President of Search and AI, told Search Engine Land that Web IQ "is used by ChatGPT for some of its web answers and it is used directly in Copilot for answers".[^9] Web IQ returns "passages and structured evidence objects" as well as documents, and Microsoft grades its results with GDSAT (grounding satisfaction), a metric covering completeness, freshness and authority.[^5] For the answers Web IQ serves, which Microsoft scopes to "some", the model can receive a passage rather than the whole page.

Citations appear inline as numbered references and as a list at the end of the response. Users can click through to source pages.

ChatGPT's research mode goes further, decomposing a query into many sub-searches and synthesising findings across a multi-step task. This is a form of agentic search, and optimising for it means being retrievable for the underlying sub-queries rather than the single original question.

## What does ChatGPT store about your page?

In most answers, ChatGPT does not open your page. It queries its retrieval pipelines, receives a title, a URL and a short snippet for each result, and writes from those three fields. Resoneo, a French SEO consultancy, captured 1,249 real ChatGPT answers in July 2026, then fetched 534 of the pages ChatGPT had cited and compared each against the snippet the index had stored for it. The stored snippet is cut just past 200 characters and taken from the start of the rendered page body, not from your meta description.[^11] The meta description still does work for results arriving via the scraped-Google pipeline, which picks it up roughly one time in three.

That snippet is anchored on the H1. Of the 463 cited pages that had one, 387 snippets contained it (83.6%), and in 77% of those the H1 was the first thing in the snippet. With a median H1 of 51 characters, that leaves around 150 characters of actual content. Whatever the template prints between the top of the body and the first paragraph spends part of the same budget: a section kicker on 29% of pages (18 characters), a publication date on 11% (25 characters), the alt text of the first image on 9% (50 characters), a site or brand name on 7% (44 characters). One page in seven had no H1 markup at all, and the snippet then starts on whichever subheading the template supplies.

The snippet is also frozen at crawl time rather than assembled per query: the same page seen under different queries returns an identical snippet, and 13% ran more than a month behind the crawl that produced them.[^11] Only OAI-SearchBot refreshes it. ChatGPT-User visits, which fetch the full page when the model decides to read one, never update what search displays, so a page can be re-read repeatedly while the stored snippet stays stale.

Two qualifications belong with these figures. Resoneo sells SEO consulting and distributes the browser extension that captured the data, and it did not test whether changing what sits above the H1 makes a page more likely to be cited.[^12] The study reads observed behaviour rather than any documented design, on a system that has already moved once during the measurement window.

## What does ChatGPT Search reward?

OpenAI has published one sentence on it: "Any public website can appear in ChatGPT search", provided OAI-SearchBot is not blocked.[^16] It names no citation signals, and nothing in its documentation mentions recency, authority, structure or headings. What can be said rests on the measurements above.

**The query, not the prompt.** ChatGPT issues its own searches, several per prompt and more specific than what the user typed; Nectiv's re-run found `official` and `gov` among the terms it adds most often.[^13] A page is retrieved for those reformulations, so the sub-questions a topic implies are what it needs to answer, not the headline question alone.

**The first 200 characters after the H1.** The stored snippet Resoneo measured is what ChatGPT writes from in most answers,[^11] so the clearest sentence on the page belongs immediately after the H1.

**Model version.** Brand-site share, fan-out depth and citation breadth all moved between GPT-5.4, 5.5 and 5.6, in both directions.[^13][^14] A pattern observed under one version is evidence about that version.

The recurring claims that ChatGPT prefers recent content, established publishers, concise passages or original analysis are observations by SEOs with no OpenAI statement and no controlled study behind them. They describe good pages, and writing to them costs nothing; treating them as ChatGPT's selection criteria is inference.

## What does ChatGPT Search deprioritise?

Nothing OpenAI has documented. Two things follow from the mechanics rather than from any stated preference. Whatever the template prints before the first paragraph spends the snippet budget Resoneo measured, so an interstitial or cookie notice whose text sits at the top of the body costs the same characters a kicker does.[^11] And a page that disallows OAI-SearchBot can still surface as "just the link and page title" if a third-party provider supplies the URL, which OpenAI says `noindex` prevents.[^16]

## How do you optimise for ChatGPT Search?

Most of the work overlaps with general GEO and SEO best practice. Specific to ChatGPT Search:

1. **Allow OAI-SearchBot in robots.txt.** OpenAI's search crawler is a separate user agent from GPTBot, which is used for training data. Blocking GPTBot does not block OAI-SearchBot. OpenAI's guidance: "For your site content to be included in summaries and snippets in ChatGPT, make sure you aren't blocking OAI-SearchBot."[^16]
2. **Maintain content recency.** For evergreen topics, update the `dateModified` schema property when you make substantive revisions. For time-sensitive topics, publish updates rather than relying on stale articles.
3. **Establish author authority.** Visible bylines, author bio pages, and consistent attribution across publications are the E-E-A-T groundwork on any surface. OpenAI has not said whether ChatGPT weights them.
4. **Protect the run-up to your H1.** The stored snippet anchors on the H1 and stops just past 200 characters, so mark up a real H1, keep kickers, dates, breadcrumbs and first-image alt text out of the space above it, and put your clearest sentence immediately after it.[^11]

## How do you measure ChatGPT Search visibility?

There is no direct equivalent of Google Search Console for ChatGPT Search. Measurement requires:

- **Manual sampling.** Periodically query ChatGPT for representative searches in your niche and check which sources are cited.
- **Referrer tracking.** OpenAI states that ChatGPT "automatically includes the UTM parameter `utm_source=chatgpt.com` in referral URLs",[^16] so filter on that parameter rather than on the `chatgpt.com` referrer alone, which an in-app browser can drop. The volume is typically small relative to organic search. As of May 2026, Google Analytics 4 added an 'AI Assistant' default channel group, with Google's documentation explicitly naming ChatGPT, Gemini, Claude, Deepseek, Copilot, and Grok as covered platforms.[^1] Perplexity is captured automatically as well, confirmed by Google Analytics in a LinkedIn post in June 2026 though not yet added to the written documentation.[^10] Google's custom channel group guide covers how to define a channel for any platform still landing in Referral.[^4]
- **Brand mention monitoring.** Tools like Mention, Brandwatch, or Google Alerts can pick up brand references in cited content.

The data is fragmentary. Treat ChatGPT Search visibility as a directional indicator rather than a precisely measurable channel.

## Frequently asked questions

### Is ChatGPT Search the same as ChatGPT plus web browsing?

Effectively yes, in current implementations. The "search" capability and the older "browse" capability have converged into a single web-retrieval feature available across ChatGPT tiers.

### Does ChatGPT Search use Google's index?

Not as its own infrastructure, but Google results do reach it indirectly. OpenAI operates its own crawler (OAI-SearchBot), and the 2026 traffic teardowns identify its default pipeline as OpenAI's in-house index.[^11] Alongside it they found results arriving through third-party data providers, and one of those pipelines scrapes Google: it is the only route on which a meta description is used, which it is for roughly a third of the results it supplies.[^11] So being absent from Google is not a clean way to stay out of ChatGPT Search.

### Do citations from ChatGPT Search affect Google rankings?

Not directly. Google has said nothing about ChatGPT citations, and there is no route by which one system's citation reaches the other's ranking. The second-order effects are the ordinary ones: a reader who arrives from ChatGPT and later searches your brand on Google, or links to you. Whether Google uses engagement as a ranking input is a separate question, covered on [does Google use clicks to rank](/seo-basics/does-google-use-clicks-to-rank/).

[^1]: [What's new in Google Analytics: AI Assistant channel — Google Analytics](https://support.google.com/analytics/answer/9164320#05132026)
[^2]: [87% of SearchGPT Citations Match Bing's Top Results — Seer Interactive](https://www.seerinteractive.com/insights/87-percent-of-searchgpt-citations-match-bings-top-results) Christina Blake and Alisa Scharf, 6 February 2025. Seer's own test: just over 500 SearchGPT citations matched against Bing and Google results for the same question; "87%+" matched Bing's top 20 organic results, against 56% for Google. Predates the 2026 retrieval-pipeline findings.
[^3]: [ChatGPT Search Is Citing Fewer Sites, Data Shows — Search Engine Journal](https://www.searchenginejournal.com/chatgpt-search-is-citing-fewer-sites-data-shows/571219/)
[^4]: [Custom channel groups — Google Analytics](https://support.google.com/analytics/answer/13051316?hl=en)
[^5]: [Announcing Microsoft Web IQ — Bing Search Blog](https://blogs.bing.com/search/June-2026/Announcing-Microsoft-Web-IQ)
[^6]: [2026 AEO/GEO Benchmarks Report — Conductor](https://www.conductor.com/academy/aeo-geo-benchmarks-report/), a live page (marked "last updated 6 July 2026" when read on 21 September 2026). The referral figures are Conductor's own platform data, US averages across its customers.
[^7]: [GPT-5.5 Instant: smarter, clearer, and more personalized — OpenAI](https://openai.com/index/gpt-5-5-instant/)
[^8]: [What can we learn from ChatGPT Search Sources? — Chris Green](https://chrisgreenseo.substack.com/p/what-can-we-learn-from-chatgpt-search), 5 July 2026: 1,000 prompts run up to ten times each, telemetry captured through Bright Data. And [How ChatGPT actually picks sources: I read the network traffic, not the outputs — Suganthan Mohanadasan](https://suganthan.com/blog/how-chatgpt-picks-sources/), 24 June 2026: about 1,240 source records from one logged-in Pro account.
[^9]: [Microsoft Web IQ](https://webiq.microsoft.ai) (Microsoft names "AI platforms like Copilot and OpenAI"), with Jordi Ribas, President of Search and AI, quoted in [Microsoft releases Web IQ, powered by Bing but designed for how AI-agents search — Search Engine Land](https://searchengineland.com/microsoft-releases-web-iq-powered-by-bing-but-designed-for-how-ai-agents-search-479194)

[^10]: [Google Analytics AI Assistant traffic measurement roll-out — Google Analytics LinkedIn](https://www.linkedin.com/posts/were-rolling-out-brand-new-automated-%F0%9D%90%80-ugcPost-7470508409834942464-SyJ8/)

[^11]: [What ChatGPT pulls, what it shows, what it cites — Resoneo](https://think.resoneo.com/chatgpt-retrieval/)

[^12]: [ChatGPT's Search Index Serves Small Sites Too, Data Shows — Search Engine Journal](https://www.searchenginejournal.com/chatgpts-search-index-serves-small-sites-too-data-shows/585388/)

[^13]: [ChatGPT Tripled Its Fan-Out Queries and Looks For Authoritative Sources — Nectiv](https://nectivdigital.com/blog/chatgpt-tripled-fan-out-queries-data-study). Roughly 4,000 prompts re-run against GPT-5.6 Sol, published 13 August 2026.

[^14]: [GPT-5.5 cites brand sites 47% of the time; GPT-5.4 did 57% — Writesonic](https://writesonic.com/blog/gpt-5-5-citation-study-vs-gpt-5-4). 50 prompts, 150 conversations, one US ChatGPT Plus account, 27 April 2026.

[^15]: [Reddit Citations Are Dropping in ChatGPT — Promptwatch](https://promptwatch.com/data/reddit-citations-are-dropping-in-chatgpt), Klaas Foppen, 18 August 2026. Promptwatch's own monitoring of the live ChatGPT, AI Overviews and AI Mode interfaces; shares are of all citations in responses that returned at least one, 7 July to 17 August 2026. The page also shows rolling seven- and thirty-day figures that update after publication.
[^16]: [Publishers and developers FAQ — OpenAI Help Center](https://help.openai.com/en/articles/12627856-publishers-and-developers-faq)
[^17]: [Overview of OpenAI Crawlers — OpenAI](https://developers.openai.com/api/docs/bots)