A training-data playbook for CPG brands: get into what AI models know before they search

Most AEO advice is about the moment an assistant searches the web. Get cited, get into the pages it reads, write answers it can lift. That work is real. But a study published on October 16 by Graphite makes a case that by the time the assistant searches, a lot has already been decided.

The argument goes like this. When you ask ChatGPT for the best high-protein snack bar, it doesn't just search for "best high-protein snack bar". It writes its own searches, often with brand names already in them. Those names come from what the model learned in training, its parametric knowledge, in the study's term. Brands the model doesn't know rarely get looked up, and when they do show up in search results, they often don't make it into the answer.

A disclosure before the numbers: Graphite sells AEO services, as Search Engine Journal noted in its coverage on October 26. The study is also observational. It doesn't control for brand awareness, which SEJ also pointed out, so famous brands may simply be famous in both training data and search results. We think the findings are useful anyway, because the steps they point to are cheap and don't depend on the study being exactly right.

What the study found

Assistants cite more often than they used to

Graphite tracked a fixed set of entity-comparison prompts in logged-out ChatGPT and Gemini. The share of responses that included a citation rose sharply for ChatGPT over two months.

Week of Aug 3Week of Sep 28

Show the data

Assistant	Week of Aug 3	Week of Sep 28
ChatGPT	56.6%	88.5%
Gemini	79%	81.2%

Source: Graphite, Parametric Knowledge Optimization, October 16, 2026. These prompts were chosen to compare entities, which tend to trigger search. Graphite cites Similarweb data showing only 7% of US ChatGPT responses across all prompt types include citations.

The note under that chart matters. Graphite's prompts are the kind that invite a search. Across everything people ask, Graphite cites Similarweb data showing only 7% of US ChatGPT responses carry citations, and SEJ cites an Ahrefs analysis that found fan-out queries in 14.4% of 66.6 million free, logged-out ChatGPT responses from August and September. Plenty of answers come straight from memory.

The searches already have brands in them

Graphite looked at about 957,000 fan-out queries generated from 22,953 non-branded prompts. 59.6% of those queries contained a brand name, and 22.7% used a site: operator to search one specific website. So for a prompt that never named a brand, the model frequently decided which brands to check before reading anything.

Being on cited pages wasn't enough

This is the finding that should worry a challenger brand. Across 3,115 entity-question pairs from 57 questions, Graphite compared how often an entity was mentioned with search off against how often it was mentioned with search on. Entities with low no-search visibility (0 to 25%) reached a median of only 11% visibility with search on, even when they appeared on 15% or more of the pages the model cited.

It also found that 51% of cited pages substantially mentioned zero or one brand, and only 25% mentioned five or more. Search results, in other words, often don't hand the model a list of brands to pick from. The list was already in its head.

This fits with what we wrote about awareness and recommendation being separate scores. Being known to people and being known to a model aren't the same, and the second takes its own work.

Why this is slow work

A model only knows what was in its training data, and that data stops at a cutoff date. Graphite calculated the gap between each model's knowledge cutoff and its release date.

Show the data

Provider	Average gap (months)
OpenAI (5 models)	4.6 mo
Google (3 models)	4.4 mo
Anthropic (4 models)	2.4 mo

Source: Graphite, Parametric Knowledge Optimization, October 16, 2026. Ranges: OpenAI 4.1 to 5.1 months, Google 3.7 to 5.1, Anthropic 1.8 to 3.0. Averages over all models Graphite tracked were longer: 8.9 months for OpenAI, 7.2 for Gemini and 3.3 for Anthropic.

Graphite calls this gap the minimum time-to-impact for new content, and SEJ points out it doesn't measure when a page is actually crawled into a training set. Either way, a review published this week shouldn't be expected in the models released over the next few months. That changes how to plan. Training data is a seasonal project, like getting into a retailer's spring reset, not a weekly optimization.

The scale is also humbling. Graphite cites an Epoch AI estimate that GPT-5 was trained on at least 30 trillion tokens. A search answer cites four to seven pages. One page in a training set is a drop; the same fact repeated across many independent pages is what moves anything.

The playbook

Step 1: Measure what the models know with search off

You can't see parametric knowledge in a normal chat window, because the assistant may search. Graphite's method is to call each model through its API with web search disabled, run each prompt 50 or more times, and count how often your brand appears. It suggests repeating the test at each model release and pinning the model version so you know which one you tested.

Write prompts the way a shopper asks, without your brand name. For a frozen dumpling brand, that might be:

• "What are good frozen dumplings for a quick weeknight dinner?"
• "Which frozen dumpling brands use no artificial preservatives?"
• "Best frozen potstickers for an air fryer"

Those are example prompts, not results. Run each one, record which brands appear and how often, and keep the file. That's your baseline. Our piece on why AI rank is the wrong metric explains why repeated runs and mention rates beat a single screenshot.

Step 2: Let the training crawlers in

Training data mostly comes from crawling. If a training crawler is blocked, the model can't learn from your pages directly. These are the main ones and what each company says they do.

Table: Training crawlers versus search and user crawlers, from each company's documentation (checked October 27, 2026)

Company	Training	Search and user requests	What the company says
OpenAI	GPTBot	OAI-SearchBot, ChatGPT-User	Each setting is independent; blocking GPTBot means content shouldn't be used for training
Google	Google-Extended (a token, not a crawler)	Googlebot	Google-Extended controls Gemini training and grounding and doesn't affect Search ranking
Anthropic	ClaudeBot	Claude-SearchBot, Claude-User	Respects robots.txt; says IP blocking may not reliably opt you out
Common Crawl	CCBot	None	Open archive used widely for training; blocked with a CCBot rule in robots.txt

Sources: OpenAI, Google, Anthropic's support center and Common Crawl.

Graphite's advice is to allow GPTBot, ClaudeBot, Google-Extended and CCBot if you want to be in training data. Check robots.txt, then check your CDN, because robots.txt isn't the only gate. Cloudflare's documentation says that since September 15, 2026, new domains block bots classified as Training or Agent by default on pages that display ads, while Search crawlers stay allowed. Its settings live under Security Settings, in the AI bot policies. If your site moved to a new domain or a new Cloudflare zone this fall, look there. Our crawler access tip and log-checking tip cover the rest.

Allowing training is a business decision, not a technical default. Some brands won't want their content in training sets. For most CPG brands, whose pages are product facts and recipes they'd like repeated, we think the trade is worth it.

Step 3: Put the facts in the HTML

Graphite recommends serving key content in server-rendered HTML. Several major crawlers don't run JavaScript, so a product page whose ingredients, sizes and claims load after the page renders may look nearly empty to them. Open your product page with JavaScript off. If the ingredient list disappears, a crawler may not see it either. The JavaScript tip has a quick test.

Step 4: Get mentioned on many different pages

Graphite notes that training data preparation removes duplicate pages and may limit how many pages come from a single domain. The same press release syndicated to 40 sites may end up counting as one page, or close to it. Graphite's advice is to aim for mentions across many unique pages and sites. In practice that means:

• Pitch reviewers and roundup writers individually instead of distributing one release widely.
• Give each retailer and partner their own description to work from rather than identical copy.
• Answer questions in your own words in communities where your category is discussed.

Each of those produces a page that says something slightly different about you, which is exactly what survives deduplication.

Step 5: Favor the sources models are known to learn from

Graphite points to three kinds of sources with a known path into training: Reddit, which has data agreements with Google and OpenAI; Wikipedia, which OpenAI has disclosed as a training source; and publishers with licensing deals, such as Axel Springer's with OpenAI.

For a CPG brand, Reddit is the practical one. A real founder account answering questions in the right subreddits, within their rules, is worth more than a dozen sponsored posts. The Reddit tip covers how. Wikipedia is harder: only pursue a page if independent coverage exists, and never write it yourself. Our Wikipedia and Wikidata tip explains the line.

Step 6: Say the same thing everywhere, with dates

A model learns your brand from many pages at once. If they disagree, it learns a blur. Graphite recommends keeping descriptions accurate and consistent and putting dates on facts that change, such as "reformulated in March 2026 with no added sugar".

That second part matters for CPG more than most categories. Products get reformulated, resized and repackaged, and the old version lives on in reviews for years. A dated statement on your own site gives the model a way to tell old from new. Our tips on one consistent entity and matching product facts are the groundwork, and our explainer on the website as a reference book for robots goes further.

Step 7: Recheck at every model release

Rerun the step 1 baseline whenever a provider ships a new model. With cutoff gaps of several months, expect changes to lag your work by at least that much, and treat a flat result after one release as normal.

The search-on side needs tracking too, since that's what shoppers see. You can run those prompts by hand in each assistant, or use a tool. XENA Intelligence is one option we know of: it asks assistants such as ChatGPT, Perplexity, Gemini and Amazon's assistant shopper-style questions and scores products on how they show up, and its SearchPanel product shows the actual answers for each SKU. Comparing the two numbers over time tells you whether memory and search are moving together.

What this playbook won't do

It won't work quickly. A model that's already released won't learn anything new about you until its successor ships.

It won't replace the search side either. When the assistant does search, the usual work on citations still decides which pages it reads and quotes. Graphite's point is that memory decides which brands it looks for, and the two feed each other.

And it rests on one vendor's observational study. The brand-in-query numbers are striking, but without controls for awareness, some of the effect may just be that big brands are big everywhere. SEJ adds a fair point: the data shows search-on and search-off visibility rising together, not that raising one will raise the other. We'd want to see the finding repeated by someone without a service to sell. Until then, the steps above are worth doing for reasons that don't depend on the study, because open crawling, plain HTML, consistent facts and real coverage help every other part of AEO too.