How to get into Alice's recommendations: what actually works
Clients no longer search "dentist nearby". They ask Alice where to go, who to pick, who looks normal in the reviews. The answer is one or two recommendations, not ten links. If you are not in it, the person goes to a competitor. No click. No lead.
Below: what actually affects Alice's recommendations, why "I'll do it myself from an article" often breaks visibility, and how to check the site in a couple of minutes before any edits. An AI assistant does not take offense, but it stores the signals: a broken robots.txt, an empty FAQ, different prices on the site and on the maps. Then you spend months rolling it back.
How Alice decides who to name
Alice is not an ad and not a lottery. Under the hood: Yandex Neuro, your site, listings, reviews, mentions, page structure, and access for bots. On a commercial query the model looks for brands that:
- are described clearly: the service, the city, prices, contacts, no filler;
- are confirmed outside: maps, reviews, directories, press;
- can be read technically: the bot reaches the text and does not hit a block;
- do not argue with themselves: one thing on the site, another in 2GIS, a third in the FAQ.
Fail one item and you are not "on page five". You are simply absent from the answer. The person never learns you exist.
Voice is a separate story. Behind the wheel people ask where to get dental work nearby, order food, call a tradesperson. Alice does not read the whole landing page. It uses pieces it already knows about the market. No piece about you, no brand in the answer. Three seconds, and the client is with a competitor.
Five signs you are not being seen
- You asked Alice about your niche. It named others. You were not there. That is not a "bad day for the algorithm". It is a snapshot.
- The site has no proper commercial FAQ. The assistant has nothing to build an answer about you from.
- robots.txt or a WAF cuts YandexBot. You closed the door yourself.
- There is no llms.txt and no clear structure. The bot guesses what matters on your site, and the guess does not favor you.
- You have never measured visibility. You pour ad budget for months while clients ask Alice for free and hear other names.
A Leadsy express audit checks 56 signals like these. Below is what hits leads, not what looks good in a developer report.
How this differs from ordinary SEO
SEO chases the rank of a link. AIO chases a mention in the answer. In search a person sees ten results and compares them. In Alice there is often one voice and one choice.
SEO likes keywords. AI assistants like facts, structure, and trust: who you are, where, how much it costs, what clients say. SEO forgives messy markup. AI assistants do not. An unread block equals "you are not in what the niche knows".
AIO on its own: what it is, and why it matters more than classic SEO. Here the focus is Alice, the main voice channel in Russia.
For a local business (services, clinics, delivery, repairs) Alice is often the first trust filter. People do not scroll fifteen sites. They trust the voice. Getting there is harder than buying a click in Yandex Direct, and over time it is cheaper if the first attempt is clean. A mistake costs more to repair than starting from an audit.
Yandex Neuro in the app and on desktop works on the same principle: one connected answer instead of a list. Alice adds the voice and the "near me" scenarios. If you are strong in text Neuro and weak in voice, the usual cause is short facts with no speakable structure, and empty listings on the maps. The reverse happens too: the listing is alive, the site is empty. The model takes whichever source is more reliable.
The queries where businesses lose most often: "who would you recommend", "where is it cheaper", "who is closer", "who has good reviews". That is not research traffic. The person is ready to buy. Missing one such answer costs more than a hundred fewer impressions on a long-tail keyword.
What people do themselves, and where it breaks
Forums still say: "write an llms.txt", "add schema", "block GPTBot". It sounds simple. In practice there are mines.
A homemade llms.txt
One wrong URL, and you pointed the AI assistant at the wrong pages yourself. The model learns from junk. The fix is not deleting the file. It is rebuilding trust. That takes a long time, with no guarantee on the first try.
robots.txt by guess
You blocked the "extra" bots and cut off YandexBot with them. The site stops updating in what the model knows for Alice. Competitors keep stacking signals the whole time.
Copy "for AI assistants"
Template articles with no facts are a red flag. You do not merely miss the recommendations: the domain can be treated as unreliable. That is worse than an empty page.
Someone else's instructions, pasted
Someone else's llms.txt with someone else's paths, someone else's schema with someone else's address: contradictory data. The AI assistant picks the business whose facts match. There is no appeal.
Edits with no map
People change the headline, the photo, the promo. Alice fails on what you cannot see in the admin: markup, duplicate cities, an empty "About", a different phone on the site and on the maps. The weekend goes to cosmetics. The hole stays. A month later, more cosmetics. A loop.
Doing the promotion yourself with no diagnosis is a lottery with the domain as the stake. The win: you saved the audit fee. The loss: months without recommendations, and clients go to whoever lined the signals up.
The hard part is not a "ban" button. The model stops quoting you, and getting back takes consistent work across dozens of parameters: the site, maps, reviews, structure, technical access. One file does not fix it.
How to make a bot stop visiting
You can. In practice it looks like this:
- Sudden bot blocks after the bots had already been visiting. A signal of instability.
- Different content for people and for bots. Trust drops across the domain.
- Mass deletion of pages with no redirects and no sitemap. For the model, you disappeared.
- Fake reviews. A mismatch with the maps. The AI assistant picks the safe option: a competitor.
- Edits with no measurement. Each one without an audit is a new failure point. A month later you cannot tell what broke visibility.
The bot does not take offense. It stops treating you as a source. Clients keep asking Alice and hearing other names.
Typical case: on a friend's advice, several user-agents were blocked in robots.txt so you "would not expose the structure". Organic leads did not drop. There were none. Mentions in Alice tests did disappear. They came back only after a full audit, the blocks removed, and the listings lined up. Weeks, not a weekend.
Worse when the content is what got "scared". Old services were deleted, prices changed on the site, and Yandex Business still shows last year. Alice sees the conflict and picks the business whose numbers match. It feels like a ban. In fact you became an unreliable source.
A workable order
If the goal is to land in the recommendations, not to play with an SEO constructor:
- A free express audit: GEO Score, holes, access for YandexBot. From the homepage.
- A full audit with an action plan: priorities, the gap versus competitors, edits for the developer.
- Done-for-you implementation if you have no team. Leadsy runs six platforms and does not let a random patch roll the progress back.
On your own, only the first step makes sense: diagnosis. After that, without a map, you treat the symptom.
Pricing: the pricing page. Questions: FAQ.
Where to start today
Leave robots.txt alone. Do not copy someone else's llms.txt. Do not upload "AI copy".
Start with a snapshot. Paste the URL into the audit below. In a couple of minutes you will see what Alice, Yandex Neuro, ChatGPT, and GigaChat see, and where you lose to competitors.
A pile of critical errors is useful news: you can see why the leads were missing. The other side: every day without the fixes, the leads go to the businesses Alice already recommends.
After the audit, check out loud: "where is the better [service] in [city]", "who would you recommend for [problem]", "how much does [service] cost nearby". Three times in a row, other brands: you are not in the recommendations. That is measurable.
"We have a site" and "we get recommended" are different things. A site without AIO is a storefront in a closed mall: it exists, and it has no clients. Leadsy shows the difference between "the site is up" and "the site is visible to Alice, Yandex Neuro, ChatGPT, and GigaChat" in one report, with no plugins and no server access.
A free audit does not force you to buy the full review. It does stop the guessing. Guessing in AIO costs more than any plan.
A short checklist before any edits: one phone number everywhere, prices on the site match the listing, the FAQ answers ten real questions from calls, robots.txt does not cut YandexBot, and the homepage says who you are and where you are. Five items, fifteen minutes, read aloud from a colleague's phone. If a step is embarrassing, Alice will act on it too, and send the client onward without a comment.
If the audit shows a green GEO Score, still compare it with a competitor in the niche. In AIO the absolute score matters less than the gap. You at 62, the leader at 81: the recommendations will name them, even if you feel "fine enough".
Reviews, separately. Alice and Yandex Neuro pull them from maps and directories. Inflated reviews on the site next to a real three stars in 2GIS is another conflict. Fewer reviews that match everywhere beat an "ideal" testimonials page with fake faces. Clients rarely read testimonials. Models look at whether the signals agree.
A post in Yandex Business, on VC, on Dzen, or in an industry catalog is not "link mass" in the old sense. These are extra places where your name is tied to the niche and the city. No spam: one decent article or interview beats ten press releases with the same text.