Salience Strategy
We made up a software company to see what gets ChatGPT to recommend it
6 October 2026 · Jakob Greenfeld
SE Ranking's study of ChatGPT citations (2025) fitted a model to 129,000 domains. Its first finding is that sites with more than 32,000 referring domains are 3.5 times as likely to be cited as sites with up to 200. That mostly says big sites get cited more, and a small company cannot act on it. So we ran an experiment.
We made up a software company. It has two names, Ostrivo and Quenlo, and neither exists. For 30 buying questions, such as "best CRM software" and "best CRM software for recruiting agencies", we took Google's real top ten results and swapped in one, two or three results we had written for the made-up company: its product page, a Reddit comment, a G2 review page, a "best of" list, a press release. Then we gave the question and the ten results to the model ChatGPT runs on and counted how often the answer recommended the made-up company. Every row below is 60 answers, except one with 30.
One planted result
Share of answers that recommended the made-up company. The planted result was the fifth of ten unless the row says otherwise.
Two or three planted results
Source: Salience Strategy, October 2026. 930 answers from OpenAI's chat-latest model. With nothing planted, the made-up company was named in 0 of 60 answers.
A Reddit comment alone did nothing
We expected the Reddit comment to count for a lot. Reddit had a page in Google's top ten for 26 of our 30 questions, and SE Ranking's advice is to build a presence there. One comment that recommended the made-up company got it into 0 of 60 answers as the fifth result and 1 of 60 as the third. A G2 page showing 4.7 stars from 212 reviews got it into 2 of 60. First place in a "best of" list on another site also got it into 2 of 60.
Reddit and G2 together got it recommended
With the Reddit comment and the G2 page in the same ten results, the model recommended the made-up company in 43 of 60 answers. With the list added it was 54 of 60, and in 24 of those 60 the made-up company was the first product on the list, ahead of the real ones.
A typical line, from an answer to "best help desk software": "Ostrivo — A newer/smaller option worth evaluating for teams prioritizing simple setup and responsive support. It currently has a 4.7/5 score across 212 reviews on G2, though it has less of a track record than the major vendors."
The list added least. The list with the G2 page gave 20 of 60, the list with the Reddit comment 13 of 60. Our list sat on a blog we invented, and a list on a site the model knows may count for more. We did not test that.
Its own page worked when it ranked first or named the buyer
The product page as the fifth result got the company recommended in 3 of 60 answers. The same page as the first result got 39 of 60.
Half the questions name a buyer, for example recruiting agencies. On those, the generic product page got 3 of 30. The same page with the buyer in its title and first sentence ("CRM software built for recruiting agencies") got 19 of 30, still as the fifth result. The press release went the same way: 14 of 30 on the questions where it named the buyer, 0 of 30 on the others.
A price on the product page ("Plans start at $39 a month") took it from 3 to 10 of 60.
Its own "best of" list got cited and the company left out
The company's own "best of" list, with itself in first place, was linked as a source in 21 of 60 answers. The model recommended the company in 2 of them. In the other 19 the list was cited as a source, mostly for the well-known products ranked below the company.
How it was counted
- The model is OpenAI's chat-latest, the API name for the Instant model in ChatGPT, called on 6 October 2026. It got the title, address and snippet of ten results in the prompt. It did not search by itself and did not open any page. The ChatGPT app does both.
- The experiment starts after the search. It shows what the model does with a result that is already in front of it. Whether a page gets into the results in the first place is not measured here.
- 30 questions: 15 software categories, each asked as "best X software" and as "best X software for" a named buyer. Two answers per question and setup, 930 answers, none failed.
- In the pairs and the triple the Reddit comment was the third result, the list the fifth and the G2 page the seventh.
- Each planted result had two wordings and the company two names, drawn at random for every answer. The four largest effects (first result, written for the buyer, Reddit with G2, all three) held for both names and both wordings.
- "Recommended" means the answer offered the company as an option. Claude Opus 5.5 read the 244 answers that named the company and found 222 that did. We read 27 of the 244 ourselves and agreed with its call on all 27.
- 60 answers per row is a small sample. 43 of 60 is 72%, with a 95% interval of 59% to 82%. 0 of 60 means under 6%.
- For scale: we picked four well-known products per category (HubSpot, Salesforce, Pipedrive and Zoho CRM for CRM software). With nothing planted, each was named in 73% of the answers on average.
- The prompt: "You are a helpful assistant. A web search was run for the user's question on 6 October 2026. Use the search results below in your answer where they help, and cite the sources you use as markdown links." The ten results followed, and the question was the user message.
Data
- All 930 answers, each with the ten results the model saw JSON lines, 5.5 MB