The method
How we measure, in full.
Ten engines, the same questions every month, and a rule that the questions may never contain your name. Here is the whole method, including the parts that make our own numbers look worse.
Rule one: your name never appears in the question
There are two different questions people confuse. “What is Big Fish Bait & Tackle?” tests whether an engine can read a website — every site with a homepage passes it. “Where do I buy live bait near Seal Beach?” tests whether the engine will offer you to a buyer who has not heard of you. Only the second one is worth money.
So every question we measure is written the way a customer would type it, and the brand name is filtered out of the drafted set before a single engine is called. A measurement that contains the brand name always looks excellent and always means nothing.
Rule two: never change the ruler mid-series
Each measurement is stamped with its own spec — how many questions, how many engines, how many repeats, written as 22p×9e×2r. Points taken under different specs are never plotted as if they were comparable, because a rising line that only rose because we added questions is a lie told with real data.
When the instrument does change, the affected points are dropped from the trend rather than quietly rescaled, and the report says how many were dropped and why.
Ten engines, four of which most tools skip
ChatGPT, Perplexity, Claude, Gemini and Grok cover the English-speaking side. Doubao, Qwen, DeepSeek, Kimi and Yuanbao cover the Chinese side, where a large share of Chinese-speaking buyers actually ask — and where no public ranking tool looks.
Engines fail in ways that look exactly like poor visibility: an expired model id, an account quota, or an engine simply slower than the timeout all produce a clean-looking zero. So every run records a per-engine delivery rate, and any engine that under-delivers is flagged on the report rather than averaged into the score as if it had answered.
What we change on your site
Structured data that matches what is visible on the page, an llms.txt that states plainly what you are and who you serve, answer-first paragraphs that lead with the answer instead of a brand story, and one consistent entity: same name, same address, same phone, cross-linked between your site, your Google profile and every listing that feeds an engine.
Inconsistency in that last item is the most common single reason an engine declines to recommend a real business. An engine that cannot tell whether two listings are the same company will recommend neither.
What we change everywhere else
Engines answer buying questions mostly by quoting other people's pages, so the work that moves the number is largely off your site: claiming and correcting the listings engines actually cite, earning inclusion in the roundups and local media they retrieve, and — where your buyers read in Chinese — the Chinese platforms.
We do not buy reviews, post fake ones, or ask anyone to. Beyond being against every platform's terms, a review graph that does not match a real customer base is exactly the pattern trust systems are built to catch.
What we will not claim
AI answers vary between runs and between users; nobody can guarantee a position in them, and anyone who does is selling something they cannot deliver. What can be done is raise the probability of being retrieved and cited, then measure whether it moved.
Technical fixes are typically visible within days. Being recommended depends on citations and entity consistency accumulating over weeks to months. Some questions stay structurally hard — a shop thirty kilometres from downtown will not win “near me downtown”, and no amount of optimization changes where it is.
See what the ten engines say about you today.
Run the free check →