A $2-a-Month Script Checks Whether ChatGPT Cites the Author's Site; llms.txt Doesn't Help
Once a month, the author asks ChatGPT 22 questions through the API, phrased the way a client would ask them, and counts how often the model links to his site. The 66 queries cost about 3 cents each, roughly $2 in total. A popular GEO (generative engine optimization) tip, the llms.txt file with hints for AI, doesn't work: 97% of such files went unrequested in May. Pages that answer a client's question directly, and mentions on third-party platforms, do help.
AI-processed from Habr AI; edited by Hamidun News
More and more people ask a neural network instead of a search engine, and a brand that is missing from its answer effectively doesn't exist for that person. According to Mediascope, in September 2026, 55% of Russians used neural networks at least once a month, compared with 26% in October 2025. A separate discipline has even emerged around this: GEO, or generative engine optimization, which is SEO's counterpart for AI answers.
The author of one Habr write-up has been running a script since August that asks ChatGPT, once a month, the questions a client would ask, and checks whether it names his site. A monthly run is 66 API calls at about 3 cents each, roughly $2 in total. Along the way, he tested which trendy GEO tips don't actually work.
Why there is no "first place" in a neural network's answer
In January 2026, SparkToro and Gumshoe collected 2,961 answers from three neural networks to the same 12 questions. The same set of brands repeated less often than in one answer out of a hundred, and the same order appeared less often than in one out of a thousand. A caveat: a co-author of the study sells visibility-tracking software, and the study has not been peer-reviewed. Hence the rule the whole measurement rests on: you can only measure the share of answers in which you appear, and only across many runs.
How the measurement works
A text file holds 22 questions. Ten are written from the viewpoint of a client who doesn't know the author, and each has a page on the site with an answer; only these go into the headline figure. Ten more concern case studies and articles, and two contain the name and domain: that is a check of the connection, not a success metric. Each question is sent three times, for a total of 66 calls. The script runs via cron on the first day of the month, and the report arrives in Telegram.
The questions go not through the ChatGPT window but through the OpenAI Responses API, using the gpt-4o-mini model with the web_search tool and tool_choice: "required", so there is no answer without a web lookup. The regular chat has account memory: when the author asked a question there, ChatGPT began reading his files and correspondence. A stranger would not be shown that. The downside of the API is that its behavior is close to consumer ChatGPT but not identical, so the measurement shows the trend under equal conditions, not what a specific person will see.
Each run goes into SQLite, keyed by date, question, and run number: whether the domain appears among the links, what position it holds, whether the brand is named without a link, and whether the model went online. Answers without search are flagged: they were assembled from the model's memory and have nothing to do with GEO.
The wording of the question decides almost everything. A query like "recommend specialists" is often read by the model as a search for nearby organizations, and it returns map links. The fix is the phrase "Give links to websites." New questions first go through a trial run without recording, and only those where the model actually cites sites remain.
Spending safeguards: a run that has already been completed is not repeated or paid for again, the limit is 70 calls per launch, there is a lock against parallel launches, the script stops after three consecutive failures, and an alert goes to a messenger.
What the measurements showed and which tips don't work
The full set ran twice, on October 1 and November 1. For the question about mentoring in vibe coding, the author appears in all three answers in every measurement, although his position ranges from first to sixth: the topic is not saturated, and the site has a page with a direct answer. In the topic of AI agents for small business the picture is the reverse: of the 9–11 cited sites, 8 repeat across all answers, and they are studio websites. Three runs is not many, so the author draws no conclusions from a single data point.
- llms.txt is a hint file for AI placed in the site's root. Ahrefs checked 137,210 domains: 28% have the file, and 97% of those files went unrequested in May. Google says its search does not use such files.
- schema.org markup (JSON-LD). On 1,885 pages, citations rose 2.4% in AI Mode and 2.2% in ChatGPT; both changes are statistically insignificant. It is still useful for rich snippets in regular search results.
- Quotes and statistics in the text. A 2023 paper promised visibility gains of up to 40%, but that was measured on its own search engine running GPT-3.5. A 2026 re-check on current systems found roughly zero: minus 0.33 percentage points for quotations and minus 0.28 for statistics, with confidence intervals spanning zero. The re-check's authors sell content evaluation, and the paper is not peer-reviewed.
What appears to hold up: AirOps analyzed 21,311 brand mentions and found that a neural network takes a brand from third-party sites 6.5 times more often than from the brand's own site (AirOps itself sells content tools). The domains that get cited instead of you are a ready-made list of places where you need to appear. Verifiable facts instead of adjectives work, as do pages with direct answers, prices, and terms.
On November 10, OpenAI launched Businesses in ChatGPT in beta: a free Business Hub dashboard where a business verifies its profile and sees mentions and clicks to its site. OpenAI Vice President Benjamin Shomair stated that the profile should not affect how often a business appears in answers. This is a channel for precise facts, not a way to buy placement.
What to do in practice
The manual version takes an evening: write 20 questions in a client's everyday words without naming the brand, add "Give links to websites," ask each one three times in a new chat, and calculate the share of answers that include your site. To track the trend, automate it: the Responses API, SQLite, the same questions every month, and spending safeguards in place before the first launch.
- Nov 21, 2026Nectar launches AI agent that finds influencers among 300 million profiles
- Nov 21, 2026opencode Hits 226K GitHub Stars: A Free 10-Tool AI Stack
- Nov 21, 2026Shield Hands GPT-6 Astra Work in Legacy Portals Without APIs: Agents 35–40% Faster
- Nov 21, 2026Val Town built a conference site in 3 minutes, while v0 took 15: testing six AI services
Key Questions Answered
Why there is no "first place" in a neural network's answer?
In January 2026, SparkToro and Gumshoe collected 2,961 answers from three neural networks to the same 12 questions. The same set of brands repeated less often than in one answer out of a hundred, and the same order appeared less often than in one out of a thousand. A caveat: a co-author of the study sells visibility-tracking software, and the study has not been peer-reviewed. Hence the rule the whole measurement rests on: you can only measure the share of answers in which you appear, and only across many runs.
How the measurement works?
A text file holds 22 questions. Ten are written from the viewpoint of a client who doesn't know the author, and each has a page on the site with an answer; only these go into the headline figure. Ten more concern case studies and articles, and two contain the name and domain: that is a check of the connection, not a success metric. Each question is sent three times, for a total of 66 calls. The script runs via cron on the first day of the month, and the report arrives in Telegram.
What the measurements showed and which tips don't work?
The full set ran twice, on October 1 and November 1. For the question about mentoring in vibe coding, the author appears in all three answers in every measurement, although his position ranges from first to sixth: the topic is not saturated, and the site has a page with a direct answer. In the topic of AI agents for small business the picture is the reverse: of the 9–11 cited sites, 8 repeat across all answers, and they are studio websites.
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.
