We Asked ChatGPT 100 B2B Marketing Questions. 94% of Cited Pages Were Inner Pages

We asked an OpenAI model with web search 100 questions that B2B marketers and export manufacturers really ask, and collected every page it cited. Of 325 citations, 305 (94%) pointed to a specific inner page and only 20 (6%) to a homepage. Official guidance from Google, OpenAI and others made up 42% of all citations, and 28 of the 100 questions were answered without any web search at all.

This post shares the data, the method and its limits, and what we think it means for a manufacturer or exporter deciding what to publish next. For the record: Blogo was cited zero times.

The study at a glance

Measure	Result
Questions asked	100 (50 English, 50 Chinese)
Questions that produced at least one citation	72
Questions answered with no web search	28, including 16 of the 20 how-to questions
Total citations collected	325
Inner pages vs homepages	305 inner (94%), 20 homepages (6%)
Citations to official guidance	137 (42%)
Citations to vendor product, feature, docs and help pages	90 (28%)
Citations to wordpress.org plugin-directory pages	17 (5%)
Citations to third-party "best tools" listicles	3 (1%)
Citations to Blogo	0

Who we asked about, and what we asked

We wrote the 100 questions by hand for two groups of people: 50 in English for overseas B2B marketers and 50 in Chinese for teams at export factories. Each language set covers five buying stages with 10 questions each:

• Problem, for example "Why is our B2B blog getting no traffic from Google?"
• How-to, for example "How to build a content calendar for a B2B manufacturer website"
• Solution or tool, for example "What is the best AI blog writing tool for B2B manufacturers?"
• Comparison, for example "Jasper vs Surfer SEO vs Frase for B2B blog content"
• Trust or risk, for example "Does Google penalize AI-generated content in 2026?"

None of the questions names Blogo. They are the questions a buyer types before knowing which vendors exist.

Finding 1: inner pages, not homepages, get cited

Out of 325 citations, 305 went to an inner page: a help article, a documentation page, a feature page, a plugin listing or a guide. Only 20 went to a homepage, and all 20 came from tool and comparison questions (11 and 9 respectively). Problem, how-to and trust questions never produced a homepage citation in this run.

Our ChatGPT web spot-checks showed the same pattern. Across the 8 unpersonalized chats we kept, ChatGPT showed 21 cited links, and one of them was a homepage.

The practical reading is simple. A model answering a specific question looks for the page that answers that specific question. A homepage that summarises everything a company does rarely is that page.

Finding 2: official guidance dominates problem and trust questions

137 of the 325 citations (42%) went to what we classed as official guidance: Google Search Central and Search Console help pages, OpenAI help and developer docs, Bing, WordPress documentation, and government sites such as nist.gov and copyright.gov.

The share depends heavily on the type of question:

Stage	Citations	Official guidance	Questions with no web search
Problem	66	60	1 of 20
How-to	19	16	16 of 20
Solution or tool	100	16	0 of 20
Comparison	102	12	3 of 20
Trust or risk	38	33	8 of 20

When a marketer asks why traffic is flat or whether Google penalises AI-written content, the answer leans on Google's own pages. The single most-cited URL in the whole run was Google's guide to creating helpful, reliable, people-first content (17 citations), followed by Google's guidance on using generative AI content (10 citations).

Vendors and agencies were almost never cited at these stages. If your plan is to win "is AI content risky?" questions with your own blog post, you are competing with the primary source.

Finding 3: most how-to questions trigger no search at all

28 of the 100 questions produced no citations because the model did not search. 16 of those were how-to questions, out of 20 how-to questions in total. Examples include "How to write a buyer's guide for a custom manufactured product" and "How to plan internal links between blog posts and product pages on WordPress".

In those cases the answer came from what the model already knew. Our spot-checks in ChatGPT itself agreed: of the 8 unpersonalized chats, 3 (one English and two Chinese questions) ran no web search.

This does not mean how-to content is pointless. Those guides still serve readers who arrive from Google, and they can become part of what models learn over time. It does mean a single how-to page is unlikely to earn a live citation for a generic question on its own.

Finding 4: tool questions cite vendor help and feature pages

Solution and comparison questions produced the most citations (100 and 102). Here the cited pages were mostly vendors' own inner pages: product and feature pages (49 citations overall), docs and help articles (41), and plugin listings.

The most-cited domains across the whole run were:

Domain	Citations	Questions citing it
google.com (incl. developers. and support.)	100	37
openai.com	25	13
wordpress.org	24	8
jasper.ai	11	10
surferseo.com	10	9
writesonic.com	9	7
semrush.com	9	9
frase.io	7	6
clearscope.io	7	5
ahrefs.com	6	5
marketmuse.com	6	3

Two patterns stood out:

• Help-center articles win capability questions. When we asked ChatGPT for a tool that tells you what to write next based on your site and search data, it cited a Surfer documentation page about adding Google Search Console, a Semrush features page, an Ahrefs help article and a Clearscope support page. Each page stated a capability in plain words.
• wordpress.org listings win WordPress questions. For "AI writer that sends drafts to WordPress for review", ChatGPT cited three wordpress.org plugin pages, a help article from another vendor and one homepage. The plugins it named were small and little known, not the biggest SEO brands.

Third-party "best tools" listicles were rare in the API run (3 citations), but they did appear in the ChatGPT web spot-checks, including dupple.com and aiunpacking.com.

Finding 5: Chinese tool questions are where homepages still win

Chinese questions produced 13 of the 20 homepage citations, against 7 for English. The biggest cluster came from one question asking which Chinese AI writing tools exist for export websites: 5 of its 8 citations were homepages of small Chinese tools. A question about how to choose a foreign-trade SEO company also cited an agency homepage. That is the one place in our data where homepages were a common answer, and the cited sites were small.

A second detail: when we asked a Chinese question about AI tools for writing English export blogs in ChatGPT itself, it cited English vendor blog posts and the same English listicle it had cited for the equivalent English question. It did not name a China-specific tool.

Finding 6: the picks change from run to run

The same question does not always produce the same answer. For the "what to write next" question, the API run cited dango.sh, the MarketMuse homepage, a Semrush solutions page and a Google Search Console help article, while the ChatGPT web spot-check cited a different set of vendor docs. For "best AI blog writing tool for B2B manufacturers", one ChatGPT chat recommended Jasper; an earlier personalized chat, which we discarded, had recommended Writesonic.

So treat patterns (page type, inner vs homepage, official vs vendor) as the signal, and treat any single "winner" as noise.

What this means for a manufacturer's website

The study was about questions for marketers, not about industrial products, so apply it with care. Still, the patterns line up with what Google says publicly. Google's AI optimization guide advises against recycling what others have already said and in favour of non-commodity content with a unique point of view. It also says special files, chunking content or rewriting just for AI search are not needed. Google's AI features documentation says there are no additional requirements to appear in AI Overviews or AI Mode beyond being indexed and eligible for a snippet.

A practical checklist:

1. Give every important question its own URL. One page per product family, material, process or capability, each opening with a plain one-sentence answer.
2. Write help-style pages, not only marketing pages. Tolerances, MOQ logic, lead-time factors, certifications, inspection steps and how to send a drawing for a quote. Vendors in our data won tool questions with exactly this kind of page.
3. Publish something only you can publish. Test results, process data, failure cases, real photos of your line. Google's helpful content guidance asks: "Does the content provide original information, reporting, research, or analysis?"
4. Do not try to out-rank Google on Google topics. For SEO and AI-policy questions, link to the official page and add your own experience instead.
5. Be listed where models already look. In our data that meant directories such as wordpress.org. For a manufacturer it may mean industry associations, trade-show exhibitor lists and recognised supplier directories.
6. Check that you are not blocking the crawlers. OpenAI's crawler documentation says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers.
7. Measure what you can. Google has added generative AI performance reports to Search Console, which show impressions, pages, countries, devices and dates for AI features in Search and Discover.

Example: a maker of stainless-steel fasteners could replace one long "Products" page with separate pages for "A2 vs A4 stainless bolts for marine use", "How we test salt-spray resistance" and "Minimum order and lead time for custom fasteners". Each answers one buyer question, and each is an inner page.

Method and limitations

• Model and search. We sent each question once through OpenRouter to openai/gpt-5.6-luna (GPT-5.6 Luna) with OpenAI's native web search at the medium search-context setting. Every request was fresh, with no system prompt and no memory. A citation is a URL returned as a url_citation annotation. All 100 questions were answered with no errors, for $1.52 in API fees.
• ChatGPT web spot-checks. We repeated 8 questions (en01, en08, en21, en22, en23, zh01, zh08, zh21) in chatgpt.com using temporary, unpersonalized chats on the default model. An earlier memory-on chat and a personalized chat were discarded as biased.
• Classification. A script stripped tracking parameters, grouped URLs by registrable domain, marked homepage vs inner page, and labelled page types by URL pattern. URL-pattern labels are approximate. The "google.com" figure groups developers.google.com, support.google.com and other Google subdomains.
• Limits. This is one run on one date (5 October 2026). The API model is not necessarily the model behind chatgpt.com's default, and picks vary run to run. The 100 questions were written by us and reflect our audience, not all B2B buyers. Treat the numbers as directional.

If you want a head start on the publishing side, Blogo picks your next topic from your site and search data, researches and fact-checks the draft against sources, and sends it to WordPress as a draft for a person to approve.

FAQ

Does ChatGPT cite homepages or inner pages more often? In our run, inner pages by a wide margin: 305 of 325 citations (94%) were inner pages and 20 (6%) were homepages, all from tool and comparison questions.

Which sites did ChatGPT cite most for B2B content and SEO questions? google.com led with 100 citations across 37 questions, mostly Google Search Central and Search Console help pages, followed by openai.com (25) and wordpress.org (24). Among tool vendors, jasper.ai, surferseo.com, writesonic.com and semrush.com were cited most often.

Why did some questions get no citations? The model chose not to search and answered from what it already knew. That happened for 28 of the 100 questions, including 16 of the 20 how-to questions.

Was Blogo cited? No. Blogo was cited zero times. At the time of the study it had very few public pages, which is consistent with what the data shows about inner pages.

Will the same questions give the same citations next week? Probably not exactly. Picks changed between our API run and ChatGPT web, and between chats. The patterns (inner pages, official guidance for trust questions, vendor help pages for tool questions) are the more stable signal.