Training Data vs Live Retrieval: Two Ways AI Knows Your Brand

AI assistants know your brand in two ways: from training data, learned before the model was released and fixed at a cutoff date, and from live retrieval, where the assistant searches the web while it answers. Training shapes long-term associations and is slow to change. Retrieval can reflect a page you updated last week, as long as crawlers can reach it. Most answers about brands blend the two, so you need a plan for each.

On this page

• What is training data?
• What is live retrieval?
• How do the two compare?
• How do answers blend the two?
• How can you tell which one an answer used?
• From MentionX data: how often answers show no sources
• What should you do for training data?
• What should you do for live retrieval?
• Should you block training but allow search?
• How long do changes take to show?
• Worked example: a pricing change
• A quick example
• How to measure each route
• Common mistakes
• Checklist
• Related
• Sources

Updated 31 October 2026. Anything marked "Our view" is MentionX's opinion, not documented engine behaviour. The worked example uses a fictional company.

What is training data?

Large language models learn from huge collections of text gathered before training: public web pages, books, code, licensed data and other sources. What the model "knows" about your brand comes from how often and how consistently you were described in that data. After training, that knowledge is frozen until the next model is trained.

Providers use specific crawlers to gather web data. OpenAI, for example, documents GPTBot as the crawler used to collect content that may be used to train its foundation models (OpenAI). Anthropic documents its own crawler and how to block it (Anthropic).

What is live retrieval?

Live retrieval means the assistant searches while it answers, then uses what it finds. This is retrieval-augmented generation (Lewis et al.). Examples:

How do the two compare?

How do answers blend the two?

Even when an assistant searches, the model's training shapes how it interprets the question, which searches it runs and how it reads the results. A brand the model already associates with a category may be framed more confidently. A brand the model has never heard of can still be cited if the retrieved pages are clear and relevant. Many visibility problems, such as an old tagline, come from training data and can't be fixed by retrieval alone.

How can you tell which one an answer used?

Clues, not proof:

• Citations or source links point to retrieval.
• Very recent facts, such as a price change last month, suggest retrieval.
• Outdated facts with no sources suggest training data.
• The same answer with search off (where products allow it) shows the model's built-in knowledge.

From MentionX data: how often answers show no sources

Method. We ran read-only aggregate queries over completed MentionX runs from 22 August to 29 October 2026 (UTC): 2,154 runs across 18 projects, mostly B2B software. Engines were called through their APIs with web search or grounding enabled. Brands are anonymised. Perplexity, Copilot and DeepSeek rows come from a single day (26 October).

How to read it. Even with search enabled, many answers came back with no sources. An answer without citations isn't proof that the engine used only training, because some interfaces don't return every source. But it's a strong hint that, for a large share of answers, there was no retrieved page you could have influenced that day. What the model already believed about your brand did the work.

Mentions without sources. Across all 2,154 runs, 41.4% mentioned the tracked brand, but in 25.8% of runs the brand was mentioned with no citation about it at all, from its own site or a third-party page. Only 1.1% of runs cited a source about the brand without naming it. Engines often talk about brands without showing where the description came from.

Caveats. API runs, which can differ from consumer apps. 18 projects. Citation behaviour depends on each API's settings.

What should you do for training data?

• Keep your brand description consistent across your site, profiles, directories and press.
• Earn coverage on widely read, reputable sources over time.
• Correct wrong facts at their source so future training data is cleaner.
• Make a deliberate decision about training crawlers, and document it.

What should you do for live retrieval?

• Allow search crawlers you want, such as OAI-SearchBot and PerplexityBot, in robots.txt (Google).
• Keep priority pages indexable and in your sitemap (Google).
• Write clear, answer-first passages with current facts.
• Show update dates honestly and make substantial updates, not cosmetic ones (Google).

Should you block training but allow search?

Some publishers block training crawlers and allow search crawlers, so they can still appear in answers. OpenAI documents them separately: a site can block GPTBot and still allow OAI-SearchBot (OpenAI). Blocking training may reduce how well future models know your brand. There's no right answer; weigh content licensing concerns against long-term visibility.

How long do changes take to show?

For retrieval, a change can appear once the page has been recrawled and reindexed, which for important pages can be days to weeks. For training data, it depends on when the provider trains and releases its next model, which you can't predict or control. Plan your measurement around both: check retrieval-based engines 2–4 weeks after a change, and expect training-based knowledge to catch up over model generations.

Worked example: a pricing change

Illustrative example. Ledgerline is a fictional AP automation company, and everything below is invented.

In June, Ledgerline drops its per-user pricing and moves to per-invoice tiers. The pricing page is updated, but a two-year-old review-site profile and several blog roundups still show the old model.

Four weeks later, the team runs "How much does Ledgerline cost?" in four engines, twice each, with search on.

• Perplexity cites the new pricing page in both runs and gets the model right.
• ChatGPT cites the pricing page once and the old review profile once. One answer is right, one is wrong.
• Two other runs return no citations and describe per-user pricing, the model's older picture.

The team's fixes follow the two routes. For retrieval, they update the review profile and email the roundup authors with the change and a link. For training, they add a short dated note to the pricing page ("We moved from per-user to per-invoice pricing in June 2026") so the change itself is described in plain text, and they keep the canonical facts page current.

They accept that the uncited answers may stay wrong until future model updates, and check retrieval-based engines again in a month.

A quick example

Illustrative. A company rebrands from "Acme Analytics" to "Acme Insight". Two weeks later, Perplexity answers use the new name and cite the updated site. A chat assistant with search off still says "Acme Analytics" because its training data predates the rebrand. The team keeps the redirect and old-name references in an "About" section, so both kinds of answer connect the names.

How to measure each route

Track the two routes separately, or you won't know which work paid off.

Before you start, write down which engines your buyers actually use. A route that matters for one engine may barely matter for another, as the table above shows.

Retrieval. For each answer, log whether it cited sources and which ones. Watch three things over time: whether your own pages are cited, whether the third-party pages you've worked on are cited, and whether answers that cite sources get your facts right. Changes here should appear within weeks of a fix, once pages are recrawled.

Training. Log answers that have no citations separately. Track whether they name you, how they describe you and whether their facts are current. Expect slow movement. When a provider releases a new model, re-run your prompts soon after, because that's when training-based answers are most likely to shift.

Logs. Your server logs show which crawlers visit which pages. Search crawler visits to updated pages are a leading signal for retrieval. Training crawler visits tell you your content is being collected, but not when or whether it will be used.

Our view: if your uncited answers are accurate and your cited answers are wrong, the problem is almost always a third-party page. If it's the other way round, your brand's long-term web footprint needs work.

Common mistakes

• Blocking every AI crawler by default. A blanket block can remove you from search answers as well as training. Decide per crawler, using each provider's documentation.
• Assuming a fix is live once the page is updated. Retrieval needs a recrawl. Check your logs for the search crawler visiting the updated page.
• Fixing only your own site. Third-party pages that repeat old facts keep feeding both routes.
• Deleting old names and facts entirely. After a rebrand or a pricing change, keep a short, dated note explaining what changed. It helps engines connect the old description to the new one, instead of treating them as two different companies or two conflicting facts.
• Testing only with search on. You'll miss what the model believes on its own, which is what many uncited answers reflect.

• Expecting training to change on a schedule. Providers don't publish when your updated pages will reach a model, so don't promise leadership a date.

Our view: treat retrieval as your short-term lever and training as your long-term reputation. Most quarterly goals should be about retrieval. Training-related work, such as consistent descriptions and earned coverage, belongs in an annual plan.

Checklist

• Search crawlers allowed
• Training-crawler decision documented
• Consistent brand description everywhere
• Priority pages current and indexable
• Old names and facts connected to new ones
• Answers checked with and without search where possible

Related

• How AI search engines build an answer
• Entity SEO for AI

Sources

• OpenAI, Overview of OpenAI crawlers
• Anthropic, Does Anthropic crawl data from the web?
• Perplexity, Perplexity crawlers
• Google Search Central, AI features and your website
• Microsoft Bing, Which crawlers does Bing use?
• Google Search Central, Introduction to robots.txt
• Google Search Central, Sitemaps overview
• Google Search Central, Creating helpful, reliable, people-first content
• Lewis et al., Retrieval-Augmented Generation