ChatGPT recommends your competitors when, for the buyer's exact question, it finds one of their pages that answers it better than yours: almost never because it has judged them better. In our test on 28 October 2026, across 8 answers to Italian B2B buying questions, 39 of 41 recommendations were backed by a page on the recommended company's own website, and only one answer in 8 leaned on comparison articles written by third parties.
So the useful question is not "why does it prefer them?". It is which page they have that you don't.
We have seen the same pattern in classic search for years. Macropix, an Italian maker of custom LED screens, went from nowhere to first place for "ledwall" (4,400 searches a month in Italy) with one page per product line instead of a single "Products" page (the full story is in our guide to SEO for manufacturers).
Sunscape, an Italian developer of solar and agrivoltaic projects, went from 329 to 5,809 organic impressions a month in six months from March 2026 (the story is in how long SEO takes).
What does Google have to do with ChatGPT?
Quite a lot. Before recommending a supplier, ChatGPT searches the web and reads pages. It picks the ones that state the service, the sector and the area plainly.
The two mistakes we see most often:
- asking the question once and treating that answer as a ranking;
- blaming reviews or the competitor's size, without opening the links under the answer.
Below you'll find the test on 7 B2B questions and 41 recommendations, the four reasons we saw, what doesn't explain the gap and a five-step diagnosis you can run on your own market. The wider measurement picture is in our AI visibility guide. Let's start.
Why does ChatGPT recommend my competitors and not me? (the short answer)
ChatGPT recommends a company when, in the pages it reads for that question, it finds proof that the company does exactly what was asked. If the proof sits on your competitor's site and not on yours, their name comes out.
A recommendation comes either from the model's memory (brands that were all over the web before training) or from a live web search, which is where the links under the answer come from.
For supplier questions ("can you recommend a company for…") the second almost always applies. And there's a technical detail few people check: according to OpenAI's crawler documentation, sites that block OAI-SearchBot won't appear in ChatGPT search answers, except as plain navigational links. GPTBot, the training crawler, is a separate setting (how to check both is in our guide to AI crawlers).
The test: 7 B2B questions, 8 answers, 41 recommended companies (28 October 2026)
We asked ChatGPT 7 questions a B2B buyer in Italy might ask, with web search on, location Italy and Italian language, through DataForSEO's ChatGPT endpoint. We ran the solar question twice on the same day. For each recommended company we noted which page ChatGPT had taken it from.
| Question to ChatGPT (Italy) | Companies recommended | Cited via their own site | Third-party sources in the answer | What they had in common |
|---|---|---|---|---|
| Solar panels for industrial sheds in Lombardy (run 1) | 6 | 6 | 0 | A page on "industrial solar" or "solar for businesses in Lombardy" |
| Same question (run 2) | 7 | 7 | 0 | 4 of 6 names kept, 3 new |
| CNC machining subcontractors in Veneto | 4 | 4 | 0 | Machining type and area in the page title |
| Custom cardboard boxes for e-commerce | 5 | 5 | 0 | Stated minimum order, from 10 to 500 units |
| Industrial 3D printing service in Milan | 3 | 1 | 0 | 2 names taken from the Google listing |
| ISO 9001 consultancy in Emilia-Romagna | 5 | 5 | 0 | An "ISO 9001" page naming the city |
| Technical translation of machinery manuals | 5 | 5 | 0 | A sector page on "industrial machinery" |
| ERP software for a small Italian manufacturer | 6 | 6 | 6 articles | Well-known brands and third-party comparisons |
| Total | 41 | 39 | 6, all in 1 answer of 8 |
7 answers out of 8 cited no ranking, no directory and no comparison article. The 2 recommendations without their own page were two Milan 3D printing labs, taken from their Google listings.
Doesn't everyone say your own site barely counts?
Yes, and for some questions it's true. A Ranqo study of 102,025 answers from five AI engines, collected between March and May 2026 and published on arXiv, found that only 2.9% of citations point to the brand's own website and 21% to "best X" list pages. That data is international, covers crowded categories, and the author holds equity in the company behind the tool used.
For niche supplier questions those lists often don't exist. Who has ever written "the 10 best CNC turning shops in the province of Vicenza"? When there's no third-party comparison, ChatGPT reads the suppliers' own sites and picks the ones that explain themselves best.
Reason 1: they have a page for the exact question (service, sector and area)
The most common reason is also the most mundane: the competitor has a page dedicated to what the buyer asked. Of the 33 companies cited via their own site (leaving out the ERP question), 23 were cited with a service or sector page, not the home page.
For solar in Lombardy, BP SRL was cited with its page "Impianto Fotovoltaico per Aziende in Lombardia" (solar systems for businesses in Lombardy). For technical translation, ChatGPT's "first choice" was Italtrust, cited with its page on translations for industrial machinery: the same words as the question. For 3D printing, Strato3D was cited with "Service Stampa 3D Milano | Conto Terzi B2B".
And the other 10, cited with their home page?
In 6 cases out of 10 the home page title already named the service and the area, like Mec-Point's "precision machining in Rossano Veneto". In practice the home page was doing the job of a service page, because the company does one thing.
It's the same logic that worked for Macropix on Google: every product line with its own URL and its own copy. A "Services" page listing eight processes answers none of the eight questions well. How to write pages built around buyers' questions is covered in our guide on how to rank on ChatGPT, with a test on 6 Italian queries.
Reason 2: they publish numbers ChatGPT can reuse (minimum orders, prices, years)
ChatGPT justifies every recommendation with a fact taken from the supplier's page. If your page has none, there's nothing to report.
In the answer on e-commerce boxes, all 5 companies came with a number from their own site: orders from 30 units for Packhelp, from 50 units with no plate costs for Scatolificio Bresciano, from 10 units for Damada, 200 and 500 units as indicative minimums for Salco, delivery in 5 to 7 working days for Imballaggi Point.
It then built a "need, company" table from those numbers: anyone wanting very few boxes is sent to whoever wrote "10 units".
The ERP answer did the same with prices. Fluentis was the only one of the six publishing a starting price, €75 per user per month for Standard and €119 for Enterprise (the price list is on its site). The result: a "how much to budget" section built entirely on Fluentis, and Fluentis among the three names in the final recommendation, next to Zucchetti and Microsoft.
Other numbers reused: systems "from around 20 kWp to hundreds of kWp" (BP SRL), 80 languages (Italtrust), over 500 technical translators (Soget Est), in business since 1979 (Vedovato).
What if my prices depend on the project?
Publish a range, a minimum order, a lead time. An indicative number beats "bespoke quote", which tells nobody anything, a model included.
Reason 3: in categories full of known brands, third-party comparisons count
The only answer in the test that leaned on third-party articles was the ERP one: 6 comparison articles and buying guides, from Talete365, ContinDigital, Sydus, Sivaf, Maia Management and Business Engine. The six brands recommended were Zucchetti, Microsoft Dynamics 365 Business Central, SAP Business One, Odoo, TeamSystem and Fluentis.
When a category is crowded with known brands, ChatGPT first reads whoever compares them, then opens the official sites. If you're missing from those comparisons, a service page alone struggles to get in.
Victorious's Quarterly Search Report for Q3 2026 points the same way on a mostly American sample: 161 brands, about 16,500 answers across nine AI platforms. 90.5% of brands were never named in answers to their own category's questions, and the factor most linked to being named was the number of sites mentioning and linking to the brand (correlation 0.49). Google Business Profile review count and rating added nothing.
How to get into the sources AI assistants read, from trade press to comparison pages, is in our article on AI citations and brand mentions. The most cited domains on Italian queries are the subject of our analysis of AI cited sources.
Reason 4: when the question names an area, the Google listing comes in (even without a website)
When the question names a city or a region, ChatGPT can add a box of local businesses with address, phone and rating. In the test this happened in 3 answers out of 8: CNC in Veneto, 3D printing in Milan, ISO 9001 in Emilia-Romagna.
And for 3D printing in Milan 2 companies out of 3 came only from there, with no page from their own site among the sources. For a business serving a territory, a complete Google Business Profile that matches the website is a second way in.
How ChatGPT picks the businesses of a city is covered in our article on local AI visibility.
What doesn't explain the gap (reviews and size)
In the test, reviews didn't decide the order. The local boxes listed 12 businesses: 10 of 12 had 10 Google reviews or fewer, some had none, and the most reviewed had 22.
The clearest case is 3D printing. C3D had 22 reviews rated 5, Strato3D had 4. ChatGPT's written "first choice" was Strato3D, the only one with a "Service Stampa 3D Milano | Conto Terzi B2B" page. Among the ISO consultancies, ARM Consulting had the most reviewed listing (10) and came fifth of five.
Size didn't decide it either: workshops with a handful of reviews sat next to NYOX, presented for large-scale systems.
So reviews don't matter?
They matter for something else: they change how AI describes a company searched by name, as we measured in our test on AI brand sentiment.
Then there's variability. The same solar question, asked twice, produced 9 different companies: 4 kept, 2 dropped, 3 new. If you tried once and weren't there, you don't know anything yet. How much this data swings is measured in our test on AI visibility data accuracy.
Every question only your competitor's site answers is a quote request you'll never see.
In the test, 39 recommendations out of 41 rested on a supplier's own page. We ask ChatGPT and Google's AI the questions your buyers ask, open the sources behind the competitors it recommends and tell you which pages and which numbers your site is missing.
Book a strategy callHow to find out why it recommends them (a 5-step diagnosis)
The diagnosis comes from reading the sources. An hour and a spreadsheet are enough.
- Write 5 to 10 buyer questions with service, sector and area: "can you recommend a company for…", "which supplier for…". You'll find real ones in Search Console too: our guide to prompt tracking explains how to choose them.
- Ask them in a temporary chat with web search on, at least twice each and on different days.
- Open every link under your competitors' names. Note whether it's their home page, a service page, a third-party article or the Google listing.
- For each competitor, copy the sentence ChatGPT uses to justify it. It almost always contains a number or a phrase taken from that page.
- Compare it with your equivalent page. If you don't have one, you've found the cause.
| What you find under the competitor's name | Likely cause | What to do |
|---|---|---|
| One of their service or sector pages | They have the page you don't | One page per service, sector and area |
| One of their numbers in the justification (minimum order, price, years) | They state facts you don't | Publish quantities, lead times, indicative prices |
| Comparison articles or buying guides | Category with known brands | Get into comparisons and trade press |
| The Google listing with address and rating | Local question | A complete, consistent Business Profile |
| No link, just the name | Answer from the model's memory | Web presence over time, not a single page |
| You don't appear even when searching your own name | Possible technical block | Check OAI-SearchBot in robots.txt |
Repeating this every month by hand gets heavy beyond 20 questions. That's why we built ScryRank, the monitoring tool developed by Visilay: it asks the same questions every month to ChatGPT, Claude, Gemini, Perplexity and AI Overviews, compares competitors engine by engine and shows which sources carry the most weight. As of 28 October 2026 the entry plan starts at €39 a month for 50 prompts. If you'd rather we did the reading and the fixes, the service is AI monitoring.
What it changes for AI citations (and on Google)
The pages that get a supplier recommended on ChatGPT are the same ones that can end up in Google's AI Overviews and AI Mode. Google states in its documentation on AI features in Search that there are no extra technical requirements: normal search best practice applies.
A page that answers a precise question, with verifiable numbers and the area served, is a better candidate for both. It's no guarantee: asked again on the same day, the same question dropped two names and added three. But without that page your name has nowhere to start from.
While you're at it, check how Bing sees your service pages: we cover it in our guide to IndexNow and Bing Webmaster Tools. If you want to work on content to get cited, the service is GEO.
Frequently asked questions
No. Ads on ChatGPT, where they run, appear below the answer labelled as sponsored and, according to OpenAI, don't influence the text (see our piece on ChatGPT Ads). A recommendation inside the answer is earned with pages that answer better, and by letting OAI-SearchBot in.
It depends: they come in when the question is local and the business listing appears, but in our test they didn't decide the order. 10 businesses of 12 had 10 reviews or fewer, and the first choice for 3D printing in Milan had 4 against another's 22.
No. Cover the same question with your own facts: service, sector, area, quantities, lead times, indicative prices. A copy adds no new facts, and facts are what ChatGPT reports.
No. GPTBot is about model training; for answers with web search, OAI-SearchBot is what counts. According to OpenAI's documentation they are two independent settings in robots.txt.
Summary
39 recommendations out of 41 rested on the supplier's own page.
The company with a page per service, sector and area wins.
Published numbers become the reason for the recommendation.
In crowded categories, third-party comparisons count.
Reviews and size didn't decide the order.
One last thing. For years a B2B company could keep a list of services on its site and leave the details to the sales team, because buyers would phone anyway. Now part of the first shortlist is drawn up by an assistant reading pages on the buyer's behalf, and whatever you haven't written down, as far as it's concerned, you don't do.
Until next time!
