Do AI assistants cite a new site in its first month?
According to Kaufman Lab data as of September 2026, AI assistants did not name the new site a single time in its first two weeks: 0 of 141 checks across two measurements. Over the same period Yandex, Russia's largest search engine, indexed 18 of its 21 pages, Google 11 and Bing none. Clicks come only from Yandex, and all 122 impressions there were searches for the lab's name, not for the topic.
- Sample
- 21 site pages, launched August 26, 2026 · two measurements, Aug 26 and Sep 9: 24 questions × 3 systems; Yandex's generative answer was not recorded · index status in three search engines for every URL, six snapshots Sep 3–9 · impressions and clicks for the week to Sep 9 from Yandex Webmaster, Search Console, Bing Webmaster Tools and Yandex Metrica
- So what
- The technical side is not the bottleneck: pages pass validation, the index grows (from 11 to 18 Yandex pages in one day), and the webmaster consoles show zero errors. The bottleneck is domain age and external mentions. Answers name trigub.ru, Dzen (Yandex's former blog platform) and RBC (a Russian business news outlet) instead of the new site, and search brings only people who already know the name.
Published Sep 2026 · data as of Sep 9, 2026 · collected by 2 scripts: one queries AI models through an API with web search, the other polls three webmaster consoles · Yandex's generative answer was not recorded in these measurements · reviewed by a person
What to do in a new site's first month: seven steps
- Finish the technical setup. Zero errors in Yandex Webmaster and Google Search Console, a sitemap submitted to both, every page loading. Do it once and leave it alone.
- Wait for indexing, not citations. AI assistants will not name a site that is not in search results.
- Record the zero. Ask AI assistants about your topic before you start and write down who they name instead of you.
- Mentions before pages. One post on a platform AI assistants already read does more than another page on your own site.
- Platforms. Habr, vc.ru, Dzen; every post links back to the original on your site.
- Track search engines separately. In two weeks Yandex indexed 18 of 21 pages, while Google moved more slowly over the same period.
- Repeat the measurement. Every two weeks. One data point means nothing; the change between points is what matters.
Which step for which situation
- The site has just launched — indexing first, citations later.
- The site is already in search results — start building mentions on third-party platforms.
- Narrow topic — two or three platforms where your readers actually spend time are enough.
Below is the measurement itself: 141 checks of AI answers over two weeks and what the webmaster consoles showed.
What we found: a new site's first two weeks in search and AI answers
AI assistants do not see a site that is not in search results. In two measurements two weeks apart, we put 24 questions about the lab's line of work to systems with web search. None of the 141 answers named the site or linked to it. In the second measurement the answers named 56 other domains instead, most often trigub.ru (3 answers), followed by dzen.ru, rbc.ru, vc.ru and habr.com with two each. A zero at the starting point is expected and useful: later movement is measured against it.
Search engines move at different speeds. In two weeks Yandex indexed 18 of 21 pages and now sends 49 clicks a week at an average position of 2. Google has 11 pages indexed but showed the site 8 times in a week, with no clicks. By September 9 Bing had no pages in its index and had crawled two.
Search traffic is still branded. All 122 Yandex impressions in the week contain the lab's name: these visitors had already heard of it on social media or through links. Topical queries, where the site has to rank on its own merits, have not appeared yet. Winning those is the work of the coming months.
What we tested: four hypotheses stated before launch
The question list, the set of systems and the counting rule were committed to the repository before the first run. A mention counts if the answer contains the domain or the Cyrillic name. The Latin-script “Kaufman Lab” does not count: a neuroscience lab in Chicago has the same name, and models confuse the two. GigaChat, Sber's AI assistant, was dropped from the measurement on August 31: it has no web search, so its zero would measure its training data, not visibility. Yandex's generative answer was not recorded in either measurement, so each had 72 planned checks; the counts are 70 and 71, as three answers were lost to request errors.
How we counted: two scripts, three webmaster consoles and a sum check
The measurement script sends 24 questions to three systems through one API with web search turned on and saves the raw answers to a public file, so anyone can recount them. Yandex's generative answer has no public API and can only be recorded by hand; it was not recorded in these two measurements. The second script polls three webmaster consoles once a day for each of the 21 URLs and records one of three states: indexed, not indexed, unknown. A pre-publish check confirms that the chart points add up to the table column.
| Lab practice | What it gave this study | Canonical name · source |
|---|---|---|
| Questions and counting rule fixed before the run | A zero cannot be explained by picking questions after the result | PreregistrationNosek et al., PNAS, 2018 |
| Raw answers are public, so the count can be rerun | Anyone can recheck 0 of 141 from the files | Data reproducibilityPeng, Science, 2011 |
| Three index states instead of two | “Unknown” does not turn into “no” | Explicit missing-data categoryLittle & Rubin, “Statistical Analysis with Missing Data”, 2002 |
| The same measurement repeated over time | Shows movement, not a single point | Panel studyMenard, “Longitudinal Research”, 2002 |
| Method changes recorded in a log | GigaChat dropped on Aug 31; Yandex's generative answer not recorded: 72 planned per run, counted 70 and 71 | Amendment protocolthe lab's own rule |
Results: AI answers, indexing, search
What to trust: two measurements, two weeks
- Sample
- One site, 21 pages, two weeks after launch. This describes one case, not statistics on new sites. There are two AI measurements: August 26 and September 9.
- Variance
- One run per question. Models with web search answer differently from run to run, and the zero is stable only because the site is not in any system's search results. Two consecutive measurements produced it independently. Non-zero values will need three runs.
- Does not generalize
- Over six snapshots the Yandex index count went 8, 8, 8, 7, 11, 18: the webmaster console shows a sample, not the full list. A count below 21 may be a reporting delay or pages actually dropping out.
- Not measured
- Yandex's generative answer was not recorded in either measurement. GigaChat was dropped on August 31: it has no web search, so its zero would measure training data. Each measurement had 72 planned checks; the counts are 70 and 71, as three answers were lost to request errors. The result is zero under any count. Rankings for topical queries were not tracked: there are only a handful so far.
- Branded impressions
- Queries that contain the lab's name, including ones typed with the wrong keyboard layout. In the week to September 9, all 122 Yandex impressions were such queries.
What to do about it: mentions before pages
The technical side is done: zero errors in the webmaster consoles, and the index grows on its own. The first mentions in AI answers will come from platforms the models already cite: Habr (a Russian tech community), vc.ru (a Russian business media site) and Dzen. Every post there links back to the original on the lab's own site. Measurements are on hold after the second run; a new version of this study will follow when they resume.
A zero at the start is unavoidable: AI assistants cannot name a page that is not yet in search results. What makes it useful is the list of sites named instead of us.
Common questions about a new site in search and AI answers
What is an AI visibility measurement for a website?
An AI visibility measurement is a regular check of whether ChatGPT, Claude and Perplexity name a specific site among their sources for a fixed list of questions. Kaufman Lab asks the same 24 questions to three systems in each measurement, 72 checks planned. Two answers were lost to request errors in the first measurement and one to a provider error in the second, so the counts are 70 and 71. Yandex's generative answer was not recorded in these measurements. Mentions of the domain and direct links are counted. The first measurement on August 26, 2026 gave 0 of 70, the second on September 9, 2026 gave 0 of 71, 0 of 141 in total; the raw answers are public.
How many days until a new site appears in ChatGPT and Perplexity answers?
In a Visiobrand study of 317,045 citations, Russian platforms picked up a new page in 7 days on average, Perplexity in 13 and ChatGPT in 18. The condition is that the page must be in the search results the systems draw their sources from. As of September 9 the new site in this study is in the Yandex and Google indexes but has not yet been named in a single answer.
Why does Yandex index a new site faster than Google?
In this case Yandex indexed 18 of 21 pages in two weeks and delivered 49 clicks at an average position of 2, while Google indexed 11 pages and showed the site 8 times with no clicks. The difference: Yandex gets signals right away from Yandex Metrica, its web analytics service, and from branded queries, while for the first weeks Google rarely shows a new domain, until external links appear: its average position is also 2, but it had 15 times fewer impressions.
What should I do if Bing does not index my site?
In two weeks Bing kept none of the 21 pages in its index and crawled two, without reporting any errors. The usual steps: verify the site in Bing Webmaster Tools, submit the sitemap and send new URLs through IndexNow. For a Russian-language site Bing brings only a handful of visits, but Copilot and some of ChatGPT's search answers use its index, so it matters for AI visibility.
How much search traffic does a new site get in its first month?
In this measurement, 422 of 2,464 pageviews in a week, or 17%. Links brought 633 views, social media 606, direct visits 579 and recommendation feeds 183. Nearly all search traffic comes from Yandex, and all 122 impressions there include the lab's name: in the first month search works as a directory for people who already know the site.
Should I check visibility in GigaChat and Alice?
Alice, yes: Yandex's generative answer draws its sources from Yandex search results and has no public API, so it can only be recorded by hand. It was not recorded in the two measurements in this study. GigaChat cannot be checked with this method: it has no web search, neither through the aggregator nor in its own API, and the model itself says it has no internet access. Its zero would show whether the domain is in its training data, not how visible the site is.
How do you count mentions of a site in AI answers?
A mention is the domain or the Cyrillic name in the answer text; the Latin spelling does not count if another well-known organization uses it. Direct links are counted separately. Domains of the systems themselves, search engines and reference sites are excluded from the list of who gets named instead. The question list and counting rule are fixed before the first run, and changes are recorded in the log.
- Kaufman Lab. Visibility dashboard: measurement 01 of August 26, 2026, raw answers, index by day.
- Yandex Webmaster. Pages in search, search query monitoring, snapshots September 3–9, 2026. webmaster.yandex.ru
- Google Search Console. Indexing and performance reports, snapshots September 3–9, 2026. search.google.com/search-console
- Bing Webmaster Tools. URL inspection and crawl, snapshots September 3–9, 2026. bing.com/webmasters
- Yandex Metrica. Traffic sources, visits and pageviews for the week to September 9, 2026. metrika.yandex.ru
- Polza. Model aggregator with web search: token and search pricing. polza.ai
- Visiobrand. Как нейросети выбирают источники: 317 045 цитат, 9 платформ [How AI assistants choose sources: 317,045 citations, 9 platforms; in Russian]. August 20, 2026. visiobrand.io
- Aggarwal P., Murahari V., Rajpurohit T., Kalyan A., Narasimhan K., Deshpande A. GEO: Generative Engine Optimization. KDD, 2024. arXiv:2311.09735
- Peng R. D. Reproducible Research in Computational Science. Science, 334(6060), 2011. doi:10.1126/science.1213847
- Nosek B. A., Ebersole C. R., DeHaven A. C., Mellor D. T. The preregistration revolution. PNAS, 115(11), 2018. doi:10.1073/pnas.1708274114