AI answers cited five articles out of 53: what I set up on kononov.dev for AI assistants

An agency published 53 articles in five weeks to get into AI answers, and a month later found only five of them among the cited sources. The same week brought a calculation showing that ChatGPT sends websites less than 1% of Google's traffic, and a Pew survey in which only 8% of Americans had asked chatbots about the elections. For a business owner being offered “promotion in AI” by an agency, these three numbers are worth more than any slide deck. Here is what I took from them and what I did on my own site.

The measurement: a third by domain, five by article

Cossa, October 23, published a study by Dmitry Zimin of the Zvonko agency. From July 2 to August 8, 2026, they published 53 articles on eight platforms and then took a measurement: 20 questions from the niche, six AI channels with search turned on, three runs per question, 1,887 source links per pass in total. The first count was by domain, and it looked good: 559 of the 1,887 links led to the platforms where they had published. But a model cites a specific page, and when counted by article identifier, 14 links and five articles remained. All five were under a month old, 12 to 26 days; the 28 articles older than 30 days were never cited. Half of the links came from one piece whose headline matched the wording of the question.

Two days earlier, Cossa, October 21, published a Pressfeed piece on GEO, which describes exactly those expectations: a company places a few expert articles, checks queries in ChatGPT or Alice AI a month later, and doesn't see its brand. A model has no publication counter after which a brand makes it into recommendations; AI search systems take what they find on the company's site and in the media. It's a piece by a service that sells media placements, and I read it with that in mind.

How much traffic this brings and how to count it

MarTech, October 22, summarized Marie Haynes's analysis of 100 sites over 12 months and more than 145 million GA4 sessions: ChatGPT sent 0.62% of the number of Google organic search sessions, all LLMs together 0.73%. Among the 55 sites where conversions were tracked, visitors from ChatGPT converted 13.8% of the time against 3.02% for Google, but by number of conversions Google still delivered 27 times more than all LLMs combined. Growth in referrals from ChatGPT, by her data, stopped in August 2025. MarTech, October 23, added the other side: organic traffic can fall while revenue holds, because zero-click answers take the informational visits, and the buyer later arrives by direct visit or by searching for the company name. The author advises watching revenue and branded queries rather than total sessions.

Pew Research, October 23, surveyed 5,273 US adults from October 12 to 18: about half use chatbots, but 8% had looked for information about the 2026 elections in them; among people under 50 it was 11%, over 50 it was 4%. This isn't about business, but it's a rare measurement of how often people ask bots anything at all.

HubSpot, October 23, laid out how AI visibility differs from SEO: a brand mention in an answer and a link to your site as a source are two different events, and they need to be counted separately. The same day, HubSpot compared eight tools that run the same questions through ChatGPT and Perplexity on a schedule, because one question gives different answers on different days.

Why this matters to the owner of a small company

Take an equipment rental company in a city of three hundred thousand. It's offered a package of ten articles on external platforms to “get into ChatGPT.” By Zvonko's measurement, five of 53 articles worked, and those dropped out of the answers after a month; by Haynes's data, all LLM traffic is under one percent of Google's. So first find out whether customers come to you from assistants at all, and only then pay for articles.

What I did on my own site

On kononov.dev, in October 2026, I added llms.txt, a separate facts page for AI assistants, and Markdown versions of the pages. An assistant with search reads a page as text, and it's easier for it to take facts from a place where they're collected without any layout. In the brief on the site, I added a “How did you hear about us” field with an “AI assistant” option. I have no numbers on the effect, and I'm not going to invent any.

I know how programs read websites from the other side. In 2023 I built Scraper Hub for a consulting company in London, a service that collected news and articles from 120 sources, mostly from regular web pages rather than XML feeds. I wrote the collection logic in Python and the server side in Node.js, and progress was shown in the interface over WebSockets. What I remember from that work is that a program has an easier time with pages where the text and the date sit in a clear structure.

What I would do

Add a “How did you hear about us” field with an “AI assistant” option to your request form, and ask the same question on the phone. It costs nothing, and a month later it gives you a first number.

Check that your service and case pages read as text: no facts inside images and no numbers that appear only after a click. If an assistant takes anything, it takes it from there.

If you decide to measure, count by page, not by domain, and repeat the same questions for several weeks. A single run shows nothing.

Compare costs with revenue, not with mentions. By Haynes's data, visitors from ChatGPT convert more often, but there are few of them; ten articles for five citations that last a month are unlikely to pay off.

About Scraper Hub: collecting news from 120 sources for a London company and the review bots that read websites. On how I read other people's percentages: numbers from vendor case studies. Earlier I wrote about measuring brand recommendations in AI. By area: the business as a single system; on the signs that a request spreadsheet is no longer enough: when a business needs a system instead of spreadsheets. If you want to understand where your customers come from and gather that in one place: Discuss your task.