How we check what AI says about you, and where it falls short.

We first confirm your business: its website (we check that your name appears on the homepage), category, and what customers come to you for. Then we write customer-style questions of four kinds:

• Category: “best [your category] [in your city]” and natural variants. Doesn’t name you.
• Problem: a need your category solves, phrased without the category name. Doesn’t name you.
• Brand: questions about you by name, such as “Is [you] any good?” or what you charge.
• Head-to-head (full report only): you versus the competitor AI recommended most.

The free check asks 2 category, 2 problem and 1 brand question. The full report asks 3 category, 2 problem, 2 brand and 1 head-to-head. For local businesses every unbranded question includes your location. Questions that are meant not to name you are checked to make sure they don’t, and every result lists the questions we asked.

The instruction is “You are a helpful assistant. Answer the user's question. Cite sources.” We call models through OpenRouter. Some providers run their own web search; the rest use a shared search plugin. We record which mode each answer used.

AI agents in the full report are covered as follows, and always shown with these labels: Hark (Asked directly); Meta Muse (Asked directly); OpenAI dots (Via the GPT model behind it); xAI Grok Bot (Via the Grok model behind it); Instinct (Not tested — awaiting permission). Agents we ask directly are asked by our team in the agent’s own app from a standard account, memory off, a fresh thread per question. Agent answers never change the score. Agents asked by hand aren’t part of the automated full report yet.

Visibility: 40 points

For each answer to a question that doesn’t name you, we give credit based on how you appear: recommended in the top 3 1.0; recommended lower down 0.6; mentioned but not recommended 0.3; absent or warned against 0. Visibility = 40 × the average credit.

“N competitors were recommended instead of you” counts businesses recommended in at least two answers with a higher average credit than yours. One-off mentions in long lists don’t count.

Accuracy: 25 points

Accuracy = 25 × correct ÷ (correct + contradicted + outdated), over the factual claims AI made about you. Claims we can’t verify from your site are left out. With fewer than 3 checkable claims we give a neutral 12.5 and say that AI said too little about you to judge.

Sources: 15 points

Among answers to unbranded questions that cite sources, the share that cite one of yours: your website, your Google profile, or your own profile on a known directory. Sources = 15 × that share. If no answer cites anything, a neutral 7.5.

Agent access: 20 points

The free score uses the 3 free models; the full report’s score uses all 10. Score version: score_v1.

We collect the factual statements each answer makes about you (hours, prices, location, services, contact details, and so on) and check them against your homepage plus up to 6 of your own pages (about, contact, pricing, services and similar), fetched without JavaScript, and your Google Business Profile when we have one.

A claim is marked contradicted or outdated only with a quote from your site. Without a quote it is “unverifiable” and doesn’t count against you. When AI contradicts your website, either the answer relied on an old or third-party source, or your page is out of date; both are worth fixing. The free check shows up to two contradictions, the most important first (hours, prices, location, services).

• robots.txt: we evaluate your rules for each AI crawler token: GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-User, Claude-SearchBot, PerplexityBot, Perplexity-User, Google-Extended, Applebot-Extended, Meta-ExternalAgent, Meta-ExternalFetcher, HarkBot.
• Fetch: we request your homepage as ourselves (PreferlyBot/1.0 (+https://preferly.ai/bot)) and record the status, redirects and any bot challenge (for example Cloudflare, Akamai, DataDome or PerimeterX).
• Readability: are your name, contact or location, and offerings present in the HTML before JavaScript runs? Many AI crawlers don’t run JavaScript.
• Structured data: is there valid JSON-LD describing your organization, business, products or services?

• Answers come from model APIs with web search. They approximate, but aren’t identical to, the consumer apps, which add their own instructions, personalization and memory.
• Each question is asked once per model: a single run. AI answers vary from run to run.
• We read answers and match business names automatically. Unusual names or near-duplicates can be miscounted.
• The fact check only knows what your website and Google profile say. If both are out of date, we can’t tell.
• We can’t see what an agent does when it browses as a signed-in user, so some access results are reported as “cannot be determined”.

• We identify as our own bot. We never send another company’s crawler user-agent.
• We never solve, bypass or evade bot protection. If we’re challenged, we report it.
• We never submit a form, place an order, book an appointment, or enter personal data on your site.
• Every agent is named with its true coverage label. We never imply we asked an agent we covered through its underlying model or haven’t tested. See who we ask.