Russian digital agency Ainet (Айnet) has published the results of an eight-month experiment, reported by CNews on 30 October. It pitted texts written by human experts against texts from five AI tools, in both Yandex and Google. The headline is that expert content ended up with up to 2.5 times the visibility in Yandex. More interesting is the path: in both engines the AI pages started well and then faded.
How the test was set up
- Ainet made up a query, “липигидродация” (“lipihydration”), so there was no existing SERP to compete with.
- It registered 10 brand-new domains with no history, no links and no Archive.org footprint. All ten used the same plain HTML/CSS design and ran on the same server, and all went live at the same time.
- Five texts came from professional beauty experts working to one brief. The other five came from Manus, ChatGPT, Perplexity, Gemini and DeepSeek, each given one shared prompt.
- Only two staff knew about the domains, and nobody searched the query. Rankings were recorded monthly in Rush Analytics.
Yandex: experts led, AI caught up, then fell back
In the first weeks Yandex gave the top five positions to the experts, with 72.7% visibility against 27.3% for the AI texts. At four months the gap had nearly closed. Manus climbed from 10th to 1st and Gemini from 8th to 2nd, taking the AI share to 45.5%. By month eight the experts were back in positions 2–6 with 71.4% visibility and an average position of 4.0, against 8.2 for AI. The Gemini, Perplexity and ChatGPT texts dropped out of the top 10.
Google: AI led the middle of the test, then lost it
Google started almost evenly, at 50.9% AI and 49.1% experts. At four months it clearly preferred the AI texts, at 63.6% visibility, and only one expert page was left in the top five. By month eight the experts were up to 67.3%, with an average position of 3.6 against 7.4, and only one AI text held on in the top six. Gemini’s page went straight to #1, stayed there for four months and finished 10th, last. The one AI exception was Manus, which ended #1 in Yandex and #3 in Google.
What Ainet says it means
Yaroslav Dombrovsky, Ainet’s head of SEO/GEO, says search engines “do distinguish generated text from expert text, they just don’t do it straight away”. He argues that matters for GEO too, because LLMs answer from sources search engines trust. CEO Ruslan Minyazhetdinov describes cheap AI content as giving “a spike for a few months, after which the investment goes to zero”. Ainet also cites a 2025 client whose impressions tripled to 1.8m while CTR fell from 6% to 2%, leaving clicks flat.
My take
I like the design. The fake query and identical sites remove most of the usual confounders. And because nobody searched the term, behavioural signals were not what separated the pages. That makes the late reversal in both engines worth noting.
But apply the evidence threshold. This is one invented beauty query and ten pages, run by an agency that sells expert content. Five expert writers and five tools is n=5 per side. Manus beating four experts in Yandex shows how much one writer or prompt can swing it. On a query nobody searches, “visibility” is a rank-tracker figure, not demand. Treat it as a well-built anecdote, not a law.
It does fit what we see elsewhere. Our Alice AI local study found answer engines lean on sources Yandex already trusts. Alice AI now answers 44% of Yandex queries, so a page that loses its rankings after four months also loses its route into the answer. The question isn’t whether AI wrote it. It is whether the page is still chosen at month eight.