Understanding ChatGPT's New Retrieval Index for Better Citation

What changed in ChatGPT retrieval?

The reported change is that ChatGPT retrieval appears to rely on its own index and a set of vertical retrieval sources, not simply on the Google results page. LovedByAI reported on September 14, 2026 that research into a retrieval index internally called Labrador described sources for the general web, PDFs, YouTube, news, arXiv, Wikipedia, local results, finance, legal, medical, shopping and images.

That matters because the old operating assumption was simple: rank in Google, then assistants will probably find you. That assumption was always incomplete. The new reporting makes the gap clearer. A page can be visible in Google while remaining unavailable, weakly represented or less competitive in a ChatGPT retrieval path.

OpenAI's crawler documentation confirms an important part of the operating model. OAI-SearchBot is the crawler used to surface websites in ChatGPT search features. OpenAI says sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, although they can still appear as navigational links. That makes crawl access a citation prerequisite, not an optional technical tidy-up.

The reported Labrador name and architecture should be treated carefully. OpenAI's public crawler documentation explains how search crawling works, but it does not publicly document every component of a proprietary retrieval index. The useful conclusion is not that publishers should chase an internal system name. It is that they should stop using one search ranking as proof that every answer engine can retrieve and cite their work.

1. Treat Google rankings and ChatGPT retrieval as related signals, not interchangeable signals.
2. Use ChatGPT citation checks to assess whether target pages actually appear in relevant answers.
3. Keep technical access, page quality and source coverage under separate review.

Who does the retrieval change affect most?

The change affects any business that relies on discovery through category questions, recommendations, comparisons or local intent. B2B SaaS teams are exposed when buyers ask ChatGPT for the best tool in a category. Local businesses are exposed when people ask for a service in a specific city. Ecommerce brands are exposed when answers draw from product and shopping sources.

It affects publishers with a narrow content strategy most. A site built around a few high-ranking commercial pages can look healthy in a conventional SEO dashboard while lacking the detail, format or source coverage that a retrieval system needs for a specific answer. This is especially likely when the useful evidence lives in PDFs, product feeds, location pages, documentation or original research that is not connected clearly to the main site.

It also affects teams that confuse bot permissions. OpenAI distinguishes OAI-SearchBot from GPTBot and ChatGPT-User. GPTBot concerns content that may be used to improve foundation models. ChatGPT-User handles certain user-initiated visits and is not used to determine search inclusion. Only OAI-SearchBot manages the automatic crawl associated with ChatGPT search visibility.

The practical implication is uncomfortable but clear. A strong Google position is a good sign, not a guarantee. A strong ChatGPT citation position is a separate outcome. That is why Citation Share is more useful than a generic visibility claim when the business goal is being chosen inside AI answers.

1. B2B SaaS teams need coverage for category, comparison and implementation questions.
2. Local businesses need crawlable service and location evidence.
3. Ecommerce teams need complete product information and shopping-ready attributes.
4. Publishers need to check technical permissions for the crawler that affects ChatGPT search.

What does the before-and-after evidence show?

The clearest before-and-after in the September reporting is not a count of Labrador documents. It is a change in the content formats receiving citations after ChatGPT 5.6. LovedByAI reported that listicles fell from 15.77% to 7.80% of ChatGPT citations, while comparison pages fell from 9.08% to 6.17%. Those figures indicate a shift in what surfaced in the observed citation set, not a universal rule that listicles or comparisons no longer work.

The same report placed this alongside Kevin Indig's analysis of 60,000 US queries and 5.32 million result rows. In that analysis, listicles held a top-three Google position for 35.9% to 39.0% of sampled queries in January, compared with 32.0% to 33.0% in August. Position-one representation was 18.9% to 20.2% in January and 15.1% to 15.6% in August.

The responsible reading is not that a format was banned or penalized. The data supports a narrower conclusion: mass-produced listicles and comparison pages may be less dependable as a citation shortcut. A page still needs a reason to be cited. Original numbers, tested methods, primary documents, transparent criteria and current details make that reason stronger.

This is the point where Citation Engineering becomes more disciplined than content volume alone. The aim is not to produce more pages with the same commercial template. The aim is to publish evidence that can resolve a specific question better than a generic summary can.

1. Before ChatGPT 5.6: listicles accounted for 15.77% of the reported citation sample.
2. After ChatGPT 5.6: listicles accounted for 7.80% of that reported citation sample.
3. Before ChatGPT 5.6: comparison pages accounted for 9.08% of the reported citation sample.
4. After ChatGPT 5.6: comparison pages accounted for 6.17% of that reported citation sample.

How should you respond to ChatGPT retrieval changes?

Start with crawl permission. Review robots.txt and confirm that OAI-SearchBot can access the public pages you want surfaced in ChatGPT search. Do not assume that allowing GPTBot or ChatGPT-User produces the same result. OpenAI explicitly separates these user agents and their purposes.

Next, inspect the pages that matter to commercial questions. A retrieval system needs more than a title containing a category phrase. It needs a clear answer near the top, defined terms, current facts, evidence a reader can trace, and page-level context that identifies what the business does and where it operates. If the page depends on a claim, make the source visible rather than burying it behind vague language.

Then widen the asset inventory. If your category has a local, product, documentation, news or PDF dimension, make sure the underlying information is accessible and complete in the relevant format. A local provider needs useful location detail. A software company needs documentation that answers implementation questions. A retailer needs accurate product attributes. The same business may need multiple assets because ChatGPT retrieval may use more than one vertical source.

Finally, test the outcome rather than admiring the setup. Build a prompt set around the questions a buyer asks before choosing. Record whether your brand appears, whether it is cited, which page or source is used, and which competitors receive citations. This gives you an answer-level measurement rather than an indirect ranking signal.

1. Review robots.txt for OAI-SearchBot on pages intended for ChatGPT search visibility.
2. Prioritize pages that answer a specific buyer question with sourceable evidence.
3. Check category-relevant vertical assets such as local pages, product data, PDFs and documentation.
4. Track citation outcomes across a consistent set of target prompts.

Should you still invest in Google SEO?

Yes. Google SEO still matters because search demand, crawling, page quality and authority are connected. The mistake is treating Google as the only measurement surface. A page that is clear, current and well supported is more likely to help both Google search and AI retrieval, but the route from one to the other is not guaranteed.

Google performance data can still reveal useful demand and content gaps. LovedByAI also reported that site-scoped searches associated with ChatGPT appeared in Search Console query data. That observation can be used as intelligence: identify pages or concepts that an assistant appears to seek, then check whether the underlying page exists and is fit to cite. Do not treat those impressions as conventional human search performance without separating the query pattern.

The better operating model is two measurements, not one. Search performance tells you how pages perform in Google. Citation performance tells you whether the brand appears as a source in relevant AI answers. The two measures can reinforce each other, but neither replaces the other.

For Omnicite, this is the practical case for measuring Citation Share: the percentage of relevant AI answers in a category that cite a brand. Rankings got you found. Citations get you chosen.

1. Keep improving search fundamentals that make pages clear and accessible.
2. Separate Google impressions from answer-engine citation outcomes.
3. Use query evidence to find missing content, then validate that content in AI answers.
4. Measure category prompts repeatedly so movement is visible over time.

What content is most likely to earn a ChatGPT citation now?

The strongest candidate is content that answers a narrow question with evidence that cannot be copied from a generic template. That can be an original dataset, a transparent comparison methodology, current product documentation, a location-specific service explanation or a first-party technical test. The point is not novelty for its own sake. The point is giving retrieval a reliable source for a claim.

Listicles and comparison pages can still work when they contain real editorial judgment and verifiable information. The reported decline in citation share means the format alone is less persuasive. A page titled around the best option should explain the selection criteria, identify limits, show sources and update dated details. A comparison should make the distinctions useful to a buyer rather than repeat vendor marketing.

Freshness is part of this. The reported retrieval stack stores crawl and publication dates, according to LovedByAI's cited coverage. Publishers should therefore review facts that date quickly, including pricing context, integrations, documentation, location details and product availability. Updating a page is not enough if the central evidence remains stale or thin.

This is not a tactic for gaming a model. It is a publishing standard. Quality, coverage and freshness give a source more chance to be useful when an answer engine needs evidence. No publisher can responsibly promise a citation count, but every publisher can improve whether its material is crawlable, specific and worth citing.

1. Publish first-party evidence where it is relevant to the buyer question.
2. Explain methods and limits on comparisons rather than relying on format labels.
3. Review fast-changing claims on a regular schedule.
4. Make documents and supporting assets easy for crawlers and readers to understand.

What should you measure after changing your content?

Measure citations first. For each priority prompt, record whether the brand is present, whether a page is cited, which URL or source appears, and which competitors are named. This provides Citation Share and Answer Presence, two measures that show different things. Citation Share shows the percentage of relevant answers that cite you. Answer Presence shows how broadly the brand appears across the question set.

Measure volume separately. Citation Count per day can show how many citations appear over time, but it should not replace share. A growing count may still leave a brand behind competitors if the answer universe grew faster. Share of Voice adds the competitive view by showing the relative presence of named competitors.

Keep a technical access record as well. A citation decline may be a content problem, a crawler-permission problem, a missing vertical asset or a change in the questions being tested. Recording the page, crawl status, prompt and cited sources makes diagnosis possible. Without that evidence, teams tend to react to a ranking graph with edits that do not address the actual gap.

The near-term goal is not to decode every internal retrieval component. It is to make sure your best evidence can be found, understood and cited. That is the work that remains useful when search interfaces change.

1. Citation Share: percentage of relevant AI answers that cite your brand.
2. Answer Presence: breadth of brand appearance across the question universe.
3. Citation Count per day: volume of observed citations over time.
4. Share of Voice: relative citation presence against named competitors.

Sources

Source: LovedByAI

Reporting on ChatGPT retrieval, the reported Labrador index, vertical sources and before-and-after citation-format changes. LovedByAI, 2026-09-14

Source: OpenAI Developers

OpenAI documentation on OAI-SearchBot, GPTBot and ChatGPT-User, including how robots.txt settings affect ChatGPT search inclusion. OpenAI Developers, 2026-09-21