ChatGPT cited Ahrefs pages, but in 43% of answers it recommended other companies' conferences
The call center of an intermediary that buys goods from the Chinese marketplace 1688 started hearing customers say the site was recommended by ChatGPT. An Ahrefs experiment shows the risk: the model cited their conference roundups, but in 43% of those answers it named other companies' events instead of Ahrefs' own. So look not only at the links, but also at whom the model recommends. Keep services and prices in HTML, and check queries in new chats.
AI-processed from Habr AI; edited by Hamidun News
A call center asks customers where they heard about the service, and more and more often hears: "GPT recommended it." A developer who helps friends with this project decided to find out for which queries the neural network names their site, which pages it reads, and what it says about the company. The project is an intermediary that buys goods from 1688 (a large Chinese wholesale marketplace) for customers from the CIS.
The main risk was shown by an experiment by Ahrefs, a developer of SEO tools: neural networks drew information from the company's pages, but in 43% of answers that cited conference roundups, they did not mention Ahrefs' own conference and recommended others instead. In other words, a link to your site does not mean you will be the one recommended. The author's notes are useful to anyone with a website offering services: part of search is moving from results pages to chat, and you need to prepare for that.
How to check for which queries you are recommended
The author admits the first method is unreliable. He asked ChatGPT to come up with "human" queries without the site's name for which it would recommend his service. The first variant read: "I want to order a large batch of goods from 1688 for resale in Kazakhstan. Recommend reliable intermediaries with websites." In another account the site did indeed come first. But the neural network was picking wording for a site it already knew, and it is unclear whether real people would ask this way.
A more reliable approach is to work backwards. First, ask the model to describe how users search for the service in your niche, including country and city. Then ask it what criteria it used to pick competitors and which pages it looked at. After that, export frequent queries from Yandex Wordstat, Google Ads Keyword Planner, Google Search Console and Yandex Webmaster. Give the export to the neural network and ask it to group queries by customer need. Have it write plain questions, keep the region, and not invent new numbers.
Next, run the questions in new chats and record whether your site appears. Note whom it recommended, in what order, and which pages it cited. An important caveat: search engine statistics show demand in search, not the number of questions asked inside GPT. Also, do not trust the neural network's explanation of why it chose a competitor, because it may invent a reason. Open the competitor's page yourself. If it lists a minimum order, buyout cost, timelines and documents, while yours says only "individual approach," the gap is obvious.
What the bot sees on the site
The author gave GPT and Claude the same link. GPT returned the content, while Claude reported that it received bare JavaScript. The author himself notes that he was working in Work mode, so the result cannot be generalized to all modes of both services. He cites a measurement by Vercel and MERJ: for the checked bots from OpenAI, Anthropic and Perplexity, no JavaScript execution was recorded, while Googlebot and Applebot do execute it. Page crawlers and browsing agents work differently, so in each test it is useful to know exactly what tool the assistant used.
The practical conclusion for a developer: the service description, prices, terms, city and contacts must be in the initial HTML. For public pages, static generation or server-side rendering is enough, while the personal account can stay as it is. Do not hide important information behind a city selector, a button or scrolling: you see the page in full, but the tool the neural network uses to read it does not perform those actions.
Also check bot access. OpenAI and Anthropic have separate bots for training (GPTBot, ClaudeBot), for search (OAI-SearchBot, Claude-SearchBot), and for fetching a page on a user's request (ChatGPT-User, Claude-User). You can block the training bot and keep the search bot. But robots.txt may allow everything while a CDN or security plugin returns 403 or a captcha to the bot. Check your server logs. A bot visit does not mean you will be recommended.
Competitors, injections and recognizability
Given the Ahrefs story, an article on "the best intermediaries" can work for competitors even if you rank first. Distinguish three outcomes: they gave a link to your site, named your company, or recommended contacting you specifically.
There are also dirty tricks: instructions for the neural network placed directly in page text, known as prompt injection. In tests by Cornell Tech researchers, in some cases 13 words were enough to change the order of recommended products. These are results from specific tests, not a ready-made instruction that works on any GPT. Microsoft described more than 50 prompts across 31 companies hidden in "summarize with AI" buttons. How many of them actually influenced future answers cannot be determined from the buttons alone. For an honest business, it matters more to tell apart your own selection, a paid placement and an independent customer review.
Classic practices still apply. Yandex states that Alice uses pages with high search rankings, and in a study by Ashmanov and Partners, the share of sources from outside the top ten has declined over time. Look at the sources in answers on your topic: if a catalog, forum or channel keeps appearing, check whether your company is there. Create listings on 2GIS, Yandex Maps and Google Business Profile with current contacts and prices. Ask for reviews in exchange for real experience, not a gift: Google Maps prohibits such rewards. For local services, check queries both with and without the city, since the answer may depend on location.
Measure the effect by recording the query, date, model and chat mode, asking clients "where did you hear about us?", and re-checking what the neural network writes about your terms.
- Nov 21, 2026Apple Is Building an AI Agent and Making It Harder for Outsiders to Get Access: Permission Now Takes 14 Steps, Not 4
- Nov 21, 2026Developer Skips Pinecone: Searching 72,450 Code Chunks in 50 ms
- Nov 21, 2026Gemini 4 Neon shrank the TS Rust binary from 47.44 to 7.97 MB in under 10 hours
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
Key Questions Answered
How to check for which queries you are recommended?
The author admits the first method is unreliable. He asked ChatGPT to come up with "human" queries without the site's name for which it would recommend his service. The first variant read: "I want to order a large batch of goods from 1688 for resale in Kazakhstan. Recommend reliable intermediaries with websites." In another account the site did indeed come first. But the neural network was picking wording for a site it already knew, and it is unclear whether real people would ask this way.
What the bot sees on the site?
The author gave GPT and Claude the same link. GPT returned the content, while Claude reported that it received bare JavaScript. The author himself notes that he was working in Work mode, so the result cannot be generalized to all modes of both services. He cites a measurement by Vercel and MERJ: for the checked bots from OpenAI, Anthropic and Perplexity, no JavaScript execution was recorded, while Googlebot and Applebot do execute it.
What are the key details on Competitors, injections and recognizability?
Given the Ahrefs story, an article on "the best intermediaries" can work for competitors even if you rank first. Distinguish three outcomes: they gave a link to your site, named your company, or recommended contacting you specifically. There are also dirty tricks: instructions for the neural network placed directly in page text, known as prompt injection. In tests by Cornell Tech researchers, in some cases 13 words were enough to change the order of recommended products.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.