Methodology

How our reports work: what we check, how scores are computed and why you can trust them.

The nine classic channels

An AI analyst walks through public sources the way a meticulous customer or a hired consultant would: searches for the company, opens its profiles, reads the reviews. Then it scores each of the nine classic channels and builds a 90-day action plan. The tenth channel, AI assistants, is described below.

What exactly gets checked

Search results for the company name and branded queries
Social profiles: activity, presentation, posting cadence
Messaging apps and contact channels: Telegram, WhatsApp Business
Maps and directories: Google Business Profile, industry listings
Review platforms: Google Reviews, Trustpilot, G2, Yelp
Mentions in the press, blogs, forums and by influencers

For every channel we save 1–3 working proof links.

The 0 to 10 scale

0: No trace of the channel found at all
1–2: Minimal: a profile exists but is empty or abandoned
3–4: Low and fragmented: something exists but isn't maintained
5–6: Average: alive and acceptable, with room to improve
7–8: Strong: active, well-kept, delivering results
9–10: Industry-leading

The score combines presence and completeness, cadence, execution quality, official status, discoverability and how feedback is handled.

Overall rating

The average of the 10 channel scores, times 10: an average of 4.3 gives a 43% rating. No hidden weights.

Honesty rules

No invented facts: when nothing is found, the report says "not found" instead of making something up.
Every finding is grounded in public sources; the important ones come with proof links.
An empty profile doesn't count as presence: it scores 1–2 and is flagged as a risk.
Small businesses typically score 3–6 per channel. That's normal: the report exists to help you grow, not to flatter you.

The tenth channel: how we probe the models

We ask each of the 10 models two questions in two layers, with search and from memory (GigaChat answers from memory only, Perplexity always searches the web), and repeat them several times, phrased the way real users phrase them. The answers are parsed into a scorecard: whether the model names you without being given your name, what it knows about you, who it recommends and who it confuses you with.

Question by name: "What do you know about company X, and would you recommend it?" Checks whether the model knows the brand, what exactly it says, how confidently it recommends it and whom it suggests instead.
Category question: "Recommend the best companies in category Y." Your name is not in the question. Shows whether you make the list, and at what rank.

Two layers: with search and from memory

With live web search — the leading layer: The model answers with access to a live web search, exactly like a customer today: assistants search the web by default. This layer moves in months, so the report leads with it.
From memory — no search: The same question without the internet: what the model recalls from training. A slow, long-term layer that shifts over years. Shown separately so you can see the gap between what AI finds and what it remembers.

What we measure

Named without your name, the headline: The share of answers to questions about your category in which AI names you. Only questions without your name count: a question with your name guarantees a mention and inflates the number. The search layer leads; memory is shown separately.
When asked about you by name: A separate 0–100 score: how many correct facts the models know about the company and how confidently they recommend it. It matters for reputation but does not count towards share of voice.
Confusion risk: Low, medium or high: whether a model mixes you up with same-name companies and attributes their facts to you.
ChatGPT and Google as a buyer sees them: The products' own answers in your market and language: ChatGPT with search, the AI Overview above Google results, and Google's AI Mode. Up to three category questions without your name and one control question with it. It shows whether Google displays an AI block for your queries and whom it names. Full report only.
Buyer-question map: About 33 questions about your category across intents and buying stages, asked on ChatGPT, Gemini and Claude. Questions containing your name are not asked; the brand row shows "name".
Where to get listed: The sites AI answers in your category cite, with a hint on what to do there. Competitors' own sites are left out and listed separately. The action plan names these sites directly.
Change over time: A repeat report for the same site is compared with the previous one: share of voice, the by-name score, which models started or stopped naming you, new competitors and newly cited sites. With monitoring, the changes arrive by email.
The audit's tenth channel: In the full audit, the "AI assistants" channel is scored from two halves: share of voice without your name and the by-name score.
Free check: One model, one run, from memory without search. It has no ChatGPT and Google block, no "where to get listed" and no change tracking.

The model roster

ChatGPT (OpenAI): search + memory
Claude (Opus 5.5) (Anthropic): search + memory
Gemini (3.8 Flash) (Google): search + memory
Grok (4.6) (xAI): search + memory
Mistral (Medium 3.5) (Mistral): search + memory
DeepSeek (V4 Pro) (DeepSeek): search + memory
Qwen (3.8 Max) (Alibaba): search + memory
Perplexity (Sonar Pro) (Perplexity): live web search
GigaChat (3 Ultra) (Sber): memory only
Alice AI (Yandex): search + memory

We run current model versions and refresh the roster as new ones ship. Every model is queried in two layers — with a live web search and from memory; Perplexity always searches the web (search-native). The Russian assistants are probed through their own APIs. GigaChat (Sber) has no live search — memory layer only. Alice AI's search layer runs on Yandex Search API's generative answer — the engine behind Yandex's own AI answers; it is available for Russian-language reports. ChatGPT and Google answers as a buyer sees them are collected via DataForSEO in the report's market and language.