Information seeking made up 20.7% of consumer ChatGPT messages in August 2026, edging past writing at 20.6% for the first time in OpenAI's Signals series. Your boss or client has probably seen that headline and wants to know which of your pages it touches. The statistic can't answer that alone, but it does give you a sorting framework, and that is enough for a useful audit.
This guide shows how to map your informational pages to OpenAI's three information-seeking categories, judge where each page type is exposed, and write a cited, versioned statement you can defend.
What the "seeking information" figure includes and leaves out
The Search Engine Journal report on the release says OpenAI's Signals V3.0 data dictionary defines seeking information as three things: requests for facts, questions about products and services for sale, and recipe or cooking help. In August 2026 the category reached 20.7% of sampled consumer messages, just ahead of writing at 20.6%. In September, writing moved back in front, 21.4% to 20.2%.
Three boundaries matter for your audit:
- How-to advice and tutoring sit elsewhere. They fall under practical guidance, which the report describes as the largest topic every month since March 2025.
- The three categories are reported as one number. Nobody can say how much of the 20% is product questions.
- There is no source data. The report says Signals gives no information on how often ChatGPT accessed external sources or cited a site.
The data describes what people ask. It does not describe where answers come from or whether any site gets a visit.
Step 1: Build the audit sheet
Create one row per informational URL. Use five columns:
- URL.
- Intent bucket: fact, product/service, recipe/cooking, or outside the three (for example, step-by-step tutorials).
- Product-question flag: yes if the page answers a question a shopper would ask about something for sale.
- Exposure note: a one-line judgment of how easily a chat answer could replace the page's core value.
- Action: keep, differentiate, strengthen, or monitor.
Pull the URL list from your CMS or a crawl. Then filter to pages that answer a question rather than sell, navigate, or serve legal needs.
Step 2: Assign each page to one bucket
Ask what the searcher wants to walk away with.
- Fact: a discrete piece of knowledge, such as a definition, a temperature, or a comparison of standards.
- Product/service: the reader is deciding whether to buy, or asking what a product does, costs, or fits. Set the product flag too.
- Recipe: the reader wants a dish or cooking technique tied to a recipe.
Pages that teach a multi-step process belong outside the three buckets, because the report places how-to advice under practical guidance. Treat that as a working rule, not a rigorous mapping. OpenAI uses an automated classifier, and your manual labels will not match it message for message. We cover related ground in affiliate site audit: measure your google traffic risk.
When a page straddles two buckets, pick the one that matches its primary query and note the second in a comment. Do not split pages into fractions.
Step 3: Judge exposure by page type
Exposure here means how much of a page's value a chat answer could deliver. This is an inference from the category definitions, not a measured outcome. The usage data contains nothing about clicks or citations.
Fact pages with generic answers are the most exposed, because their value is a short, widely available piece of knowledge. Fact pages with original data, tables, or firsthand testing are less exposed, since their value is the evidence itself.
Product/service pages deserve the most attention because they carry commercial intent. The flag lets you check that specifications, pricing context, compatibility, and returns details are complete and consistent. Recipe pages are exposed on the basic method but keep their value through tested specifics, photos, and variations.
Step 4: Review the illustrative 20-page example
Illustrative example only. This is a hypothetical kitchenware retailer's blog and guides. The URLs, labels, and judgments are invented to show the method, and no traffic, ranking, or citation results are implied.
| # | URL | Bucket | Product flag | Exposure note | Action | |---|---|---|---|---|---| | 1 | /guides/what-is-sous-vide | Fact | No | Generic definition | Differentiate with original testing | | 2 | /guides/meat-doneness-temps | Fact | No | Short lookup table | Add source notes and tested ranges | | 3 | /guides/convection-vs-conventional | Fact | Yes | Comparison leads to purchase | Link to relevant products | | 4 | /guides/how-long-does-rice-keep | Fact | No | Single-answer query | Monitor | | 5 | /guides/cast-iron-vs-carbon-steel | Fact | Yes | Comparison with buying angle | Strengthen with side-by-side data | | 6 | /guides/oven-sizes-explained | Fact | No | Reference content | Keep, update annually | | 7 | /guides/knife-steel-glossary | Fact | No | Glossary, generic | Differentiate or consolidate | | 8 | /products/dutch-oven-5qt | Product | Yes | Specs and compatibility questions | Verify specs, add FAQs | | 9 | /products/stand-mixer-pro | Product | Yes | Features and price questions | Verify specs, add comparisons | | 10 | /reviews/best-air-fryers | Product | Yes | Heavy shopper intent | Priority: refresh testing notes | | 11 | /reviews/best-chef-knives | Product | Yes | Heavy shopper intent | Priority: refresh testing notes | | 12 | /guides/which-blender-to-buy | Product | Yes | Decision-stage content | Priority: clarify who each pick suits | | 13 | /guides/is-nonstick-safe | Product | Yes | Safety concern, trust-sensitive | Cite primary sources | | 14 | /products/espresso-machine-x | Product | Yes | Questions about fit and upkeep | Add maintenance details | | 15 | /guides/warranty-and-returns | Product | Yes | Policy questions | Keep accurate and current | | 16 | /recipes/weeknight-risotto | Recipe | No | Basic method widely available | Differentiate with tested tips | | 17 | /recipes/sourdough-starter | Recipe | No | Method plus troubleshooting | Strengthen troubleshooting | | 18 | /recipes/slow-cooker-chili | Recipe | Yes | Links to appliance | Check appliance links | | 19 | /tutorials/season-cast-iron | Outside | Yes | Practical guidance | Monitor, treat separately | | 20 | /tutorials/sharpen-a-knife | Outside | Yes | Practical guidance | Monitor, treat separately |
The tally: 7 fact pages, 8 product/service pages, 3 recipe pages, and 2 outside the three buckets. See full coverage of how to validate an affiliate marketing niche in 5 steps for additional background.
The product flag spreads wider than the bucket count suggests. Thirteen of the 20 pages carry it: all 8 product pages, 2 fact pages, 1 recipe page, and both tutorials. On this invented site, the work queue starts with rows 10 to 12, not the glossary.
Step 5: Prioritize without chasing the headline
Rank rows by two things: commercial value to your business and how replaceable the page is. A high-value, easily replaced product guide outranks a low-value glossary entry. This is a judgment call, and you own it.
Don't treat the 20% as a traffic forecast. The report does not say how many of those messages reach a website, and a share of messages is not a share of your potential audience.
Step 6: Verify and cite the figures
Before you put a number in a deck, run this checklist:
Also read: see how to label free articles on a paywalled site: test plan
- Metric name: write "share of sampled consumer messages classified as seeking information," not "share of ChatGPT searches."
- Month: name the month, since the series is monthly.
- Version: write Signals V3.0. The report says each changelog update adds a quarter of data and revises earlier months, so the same month may read differently later.
- Sample basis: 1M messages per month from accounts with a declared age of 18+ on Free, Go, Plus, or Pro plans. Enterprise and Codex are excluded, and an automated system classifies each message by subject area.
- Source link: link the release you used. The Search Engine Journal report summarizes it and covers the Oct. 26, 2026 release.
A safe sentence: "In OpenAI Signals V3.0 (released Oct. 26, 2026), information seeking made up 20.7% of sampled consumer ChatGPT messages in August 2026, based on 1M messages per month from 18+ accounts on Free, Go, Plus, or Pro plans."
Cross-check against the 2025 paper
The 2025 OpenAI and Harvard paper put information seeking at 24% of usage by mid-2025. Signals shows 18.3% for June 2025, the paper's final month. That is a gap of about 5.7 points.
According to the report, the data dictionary says both use the same classifier prompts and the same three categories. The sampling differs. The paper drew about 1.1 million conversations, one message each, and averaged shares over 28-day periods. Signals samples 1M messages monthly, and V3.0 restated earlier months to match that size.
Neither the release blog nor the data dictionary explains the gap. Cite both numbers with their sources, and do not blend or average them. If a stakeholder asks which is right, the accurate answer is that the two sources differ and the reason is unpublished.
Limitations box
What this data cannot support.
- Consumer plans only. It samples Free, Go, Plus, and Pro accounts with a declared age of 18+. Enterprise and Codex use is absent.
- Automated classification. Individual messages may be mislabeled.
- Combined categories. You cannot read off how much is product questions versus facts versus recipes.
- No source data. It says nothing about web citations, sources used, or referral traffic.
- Revised figures. Numbers change between versions, and V3.0 changed earlier months.
Do not use it to claim that ChatGPT sends or removes visits to any site.
What to expect next
Expect the numbers to move. Each Signals update adds a quarter and revises earlier months, so recheck your cited figure whenever a new version appears and keep the version label in your sheet.
Treat the audit as a living document. Relabel pages when your content changes. Add your own measurements, such as referral data from your analytics, before you decide that any page type needs a different strategy.
