A note went around this week: ChatGPT is indexing the web, the index is called Labrador, free Instant answers barely open a page, and Bing rank is not ChatGPT rank. We checked it against the public measurements — RESONEO (July corpus, August replay), Search Engine Land, Peec AI, and OpenAI’s own crawler docs. The core is true. A few lines in the note need a tighter reading.
Quick answer
ChatGPT now has its own retrieval hub, seen in traffic as labrador. It is not Bing. Researchers who ran the same queries found about 1.5% of Labrador URLs in Bing’s top 20. Free Instant mode answers from a stored title plus a ~200-character snippet and opened zero pages in 93% of captured answers. Paid Thinking is almost the opposite: about 75% scraped Google, and that is when ChatGPT-User actually reads the page. Treat Instant as an H1-and-first-paragraph contest on OpenAI’s cache. Treat Thinking as crawlable HTML plus classic Google SEO. Treat Copilot as Bing. Those are three doors, not one rank.
Best for: teams who already do SEO or “AI visibility” and have been using Bing Webmaster or a ChatGPT Plus session as a proxy for what free users see. Honest limit: Cipher did not capture the result_source stream. The measurements below are RESONEO, Peec, and Search Engine Land, dated July–September 2026. OpenAI has not published a Labrador spec. The field that named the pipeline vanished on 21 July, so later studies infer the engine from snippet shape. We translated the findings onto this site — robots.txt, blog chrome, SSR — and wrote a review list for the rest of the archive.
Last updated: 19 September 2026. Cipher Projects is an Australian-led engineering studio. We build production software and the pages that have to be legible to Google, Bing, and ChatGPT. We are not an SEO agency selling a Labrador package.
Is the Labrador note true?
Most of it. The useful work is splitting what the traffic showed from what people inferred.
| Claim in the note | What we can check | Verdict |
|---|---|---|
| ChatGPT is indexing the web; the internal name is Labrador | RESONEO and Peec read a result_source field on every web result from 21 May to 21 July 2026. Four values: labrador, bright, oxylabs, serp. OpenAI does not document the name. Peec also maps Labrador as a family of vertical indexes (web, news, PDF, YouTube, Reddit, shopping, and others). |
Confirmed as independent traffic analysis. Not an official OpenAI product name. |
| Three of the labels are Google scrapes; the fourth is OpenAI’s own index | RESONEO replayed queries: bright and oxylabs match Google (about one URL in three on Google page one; titles match including ellipsis). serp is the thin open-web / SerpAPI path. Labrador matches neither Google nor Bing snippets. |
Confirmed. |
| Only about 1.5% of Labrador results sit in Bing’s top 20 | Search Engine Land / RESONEO, same fan-outs. Extra tells: Bing caps displayed titles at 75 characters; 24% of Labrador titles run past that. Snippet reuse between Labrador and Bing is nil. | Confirmed. Bing rank is not ChatGPT Instant rank. |
| Free users mostly get Labrador; Instant opened no page in 93% of answers | Instant: Labrador for stable / business / product questions (97–100%); news ~50/50 with scraped Google. Zero pages opened in 93% of Instant answers. The model sees title + a snippet cut at 202 characters. August update: free Think still leans Labrador (74.7%); it is not the same as paid Thinking. | Confirmed, with that news-and-Think caveat. |
| Paid Thinking is ~75% scraped Google and actually reads pages | Paid Thinking, medium effort: 75.3% Google / 24.7% Labrador (August replay). Opens happen almost only here. An opened page is cited 74% of the time; a page only in the grounding set is cited 7%. | Confirmed. |
| The snippet comes from the H1 and nearby text; meta description is ignored | RESONEO compared 534 cited pages to stored snippets. Of pages with an H1, 83.6% of snippets contain it. Meta description is ignored on Labrador (Google-scrape still uses it ~1 in 3). The snippet is frozen at crawl time and does not change with the query. | Confirmed. |
| Updating a page does not update ChatGPT; unpopular pages stay stale for months; no JavaScript | Two stores: a snippet index (OAI-SearchBot) and a full-page Markdown cache (ChatGPT-User). Cache is shared across users and plans. Fresh for ~30 minutes, then stale-while-revalidate. Oncrawl saw copies served 90+ days later. Cache-Control: no-store and noindex were ignored in their tests. Fetches do not execute JavaScript; pages over 4 MB are rejected. |
Confirmed as researcher measurement. OpenAI has not confirmed the cache policy. |
| Three bots: GPTBot, OAI-SearchBot, ChatGPT-User. Blocking all three hides you from search | OpenAI crawler docs: GPTBot = training; OAI-SearchBot = ChatGPT search inclusion; ChatGPT-User = live user fetch (robots.txt may not apply). A fourth bot, OAI-AdsBot, validates ads landings. You can allow SearchBot and disallow GPTBot. | Confirmed. The note is missing AdsBot. The “blocked all three, then wondered why ChatGPT never cites us” pattern is real. |
| ChatGPT still mixes Google and Bing; shopping is already testing “prefer our own index” | Google mix is real via bright / oxylabs. A bing pipeline showed up on a Team account in June (Oncrawl); RESONEO saw none on free/Plus. Peec logged live shopping experiments on 2 September 2026, including prefer-index-over-serp-v3 and shopping-index-q2qb. |
Google mix: confirmed. Bing mix: Team-plan hypothesis, not the default free path. Shopping A/B: confirmed as Peec traffic, names will change. |
One-line standup version: OpenAI built an index, free Instant lives on it, and the record it stores of your page is your title plus about 200 characters around the H1.
What does this change for AI SEO?
It splits “show up in ChatGPT” into four leaderboards. We already treated Google rank, AI recommendation, and AI citation as three different scores. Labrador adds a retrieval door that does not inherit Bing and only half-inherits Google.
| Door | Who sees it | What the model actually has | What to optimise |
|---|---|---|---|
| Google organic / AI Overviews | Searchers on Google | The live page, plus Google’s own snippet | Classic SEO. Meta description still matters here. |
| Bing / Copilot | Copilot and some ChatGPT Team traffic | The Bing index | Bing Webmaster, IndexNow. Do not use this as a ChatGPT Instant proxy. |
| ChatGPT Instant (and free Think) | Most ChatGPT users | Labrador title + ~200-character snippet. Almost never the live page. | OAI-SearchBot allowed. H1 that stands alone. First paragraph that answers. Date and author in visible text. HTML that exists without JavaScript. |
| ChatGPT paid Thinking | Paying users who wait | ~75% scraped Google, then a live or cached Markdown fetch | Google rank plus a page ChatGPT-User can GET as HTML. Opening the page triples the citation rate in RESONEO’s set. |
The mistake the note is killing is the 2024 habit: “we rank on Bing, therefore we rank in ChatGPT.” That was a decent guess when people thought ChatGPT Search was a Bing wrapper. It is a bad operating assumption in September 2026.
Why Instant is a 200-character contest
In Instant, ChatGPT does not “read your article.” It is handed a result list the way a search engine would: URL, title, snippet. Labrador passes the <title> whole — including titles Bing would truncate — then a snippet cut just past 200 characters, taken from the start of the rendered body. Typical split on pages that have an H1: a few characters of chrome, about 50 characters of H1, about 146 characters of whatever sits next.
What eats that 146 is not your carefully written meta description. RESONEO’s losses, in order: section kicker / category (29% of pages, ~18 characters), publication date (~25), alt text of the first image under the title (~50), brand name, breadcrumb, byline. One page in seven had no H1 at all, and then the snippet starts on whatever the template printed first.
That is why “answer the question in the first paragraph” is not a style tip. It is the only paragraph Instant is likely to see. It is also why a long subtitle, a category eyebrow, and a hero image whose alt repeats the title are not free. They spend the budget before the answer starts.
The snippet does not update when you publish a fix. ChatGPT-User can refresh the full-page cache and leave the Labrador snippet untouched. SearchBot writes the snippet. If nobody is asking about the URL, both copies can sit for months. A page you “updated yesterday” can still be last summer in Instant.
What we would change on a real site
The note’s to-do list is the right list. We ran it against cipherprojects.com before writing this page.
| Move | Why | What we found here |
|---|---|---|
Allow OAI-SearchBot in robots.txt, separately from GPTBot |
Search inclusion and training are different switches. OpenAI says so. | GPTBot and ChatGPT-User were named. SearchBot was only covered by User-agent: *. We added an explicit allow. If you blocked “all OpenAI bots” to stay out of training, turn SearchBot back on. |
| Write an H1 that still makes sense as the only line someone reads | 83.6% of Labrador snippets include the H1. | Our titles are the H1. Vague titles (“A complete guide to…”) waste the 50-character block. Question-shaped or claim-shaped titles survive better. |
| Answer in the first paragraph | That is the 146 characters Instant has left after the H1. | Newer posts already open with a Quick answer box. Older posts often warm up. Those are the first tweaks on the review list. |
| Put date and author in visible text near the top | The model cannot quote a byline that exists only in JSON-LD. | Layout already prints author and date under the H1. Keep a “Last updated” line in the body so a refresh is visible in the Markdown cache too. |
| Render the article on the server | The fetch robot does not run JavaScript. Client-only copy is invisible. | Blog HTML is SSR. The risk is a future page that hydrates the body from a client fetch. Check with curl, not only the browser. |
Two site-wide notes the original list does not mention, because they only show up when you look at a real template. Category text sits above our H1, so it can spend the prefix. The framed hero uses the post title as image alt, so Labrador can spend ~50 characters repeating the H1. Those are layout bugs for Instant, not copy bugs. They are on the review checklist as a single chrome pass, not 160 individual rewrites.
Keep Bing Webmaster and IndexNow. Copilot still drinks from Bing. Paid Thinking still drinks from scraped Google. Labrador does not replace those doors. It just stops you from pretending they are the same door.
FAQ
Does this mean I can ignore Google? No. Paid Thinking is still ~75% scraped Google, and Instant still splits news with Google. Google is one corpus. It is no longer the only corpus free users see.
Does a good Bing rank help ChatGPT? Not for Instant. Overlap with Labrador’s top results was about 1.5%. It still helps Copilot, and it may help ChatGPT Team if the bing pipeline is real on that plan.
We blocked GPTBot. Are we invisible? Only if you also blocked OAI-SearchBot, or a WAF that treats all OpenAI IPs as one bot. Training opt-out and search opt-out are separate. Allow SearchBot if you want citations; keep GPTBot disallowed if you do not want training.
We updated the post this morning. Why is ChatGPT still wrong? Instant is reading a snippet frozen at the last SearchBot visit. The full-page cache refreshes only when someone asks again and the copy is older than ~30 minutes — and even then the next visitor gets the new copy, not the current one. Unasked pages can sit for months.
Should we write for Instant or for Thinking? Write the H1 and first paragraph for Instant. Write the rest of the page for Thinking and for Google. One page, two depths. Do not ship a 200-character stub and call it a guide.
Will OpenAI keep mixing Google in? They are already testing “prefer our own index” on shopping traffic. Expect more Instant-class answers to come from Labrador over time. Measure again; do not freeze a July 2026 mix as policy.
Who should do this work for an Australian or Singapore team? Whoever already owns the HTML, robots.txt, and the money URLs — not a prompt-engineering vendor. Cipher Projects will review crawl rules, SSR, and the first-paragraph answers on the pages that back a buying decision. Start at contact.
Related: AI and the indie-hacker channel shift · Best AI development companies in Australia · AI transformation and consulting · OpenAI forum / SSO / Codex · Company facts
Research credit: RESONEO (Olivier de Segonzac, with Jérôme Salomon at Oncrawl), Search Engine Land, Peec AI (Malte Landwehr, Tomek Rudzki), and OpenAI’s crawler documentation. The note that started this page was a friend’s digest of that stack.