LLM visibility tool: measure each model, not the average

Two mechanisms decide whether a model knows your brand, and models weigh them differently. Training memory: what the model absorbed before its knowledge cutoff, dominated by heavily-cited sources like Wikipedia, major publishers, and well-linked industry coverage. Live retrieval: what the model fetches from the web while answering, which reflects your current search presence instead of your historical coverage.

A retrieval-heavy model can cite an article you published last week; a training-heavy model may not register you until its next release, however good your recent content is. This is why the same brand routinely scores high on one model and near zero on another, and why a single blended "AI score" hides the information you actually need: which model is the problem.

• Per-model mention rate: your brand's appearance rate in each model's answers to the same fixed prompts, reported separately, never averaged away.
• Prompt-level answers: the actual responses, stored, so you can read what each model says about you rather than trust a number.
• Citations per model: which URLs each model draws on. Retrieval-grounded models expose these directly; they are the levers for changing the answer.
• Competitor context: who each model names instead of you, per prompt, since the same competitor rarely dominates every model.
• Change over time: model releases recut answer sets, so the tool must track across updates, not just within one model version.

The cheapest first step is a one-shot check. Meev's free checkers run five buyer-discovery prompts in your category through a model of your choice and score the result, with no signup for the score: try the ChatGPT checker, Claude, Gemini, Perplexity, or the all-model aggregate. A low score on one model and a high score on another is normal, and it is precisely the signal that tells you where to work.

Run one mid-size SaaS brand through five models on the same question and the spread is routinely dramatic: strong on two models that lean on live retrieval, because the brand ranks well for its category queries; weak on a training-memory model whose last cut predates the brand's growth; absent entirely on another that favors heavily-cited reference sources the brand never earned. One brand, one question, four different realities, and a blended score would have reported it all as "moderate visibility" and pointed at nothing.

Measured per model, each weakness names its own fix. The retrieval engines are already won, so protect them with freshness. The stale training-memory model will self-correct at its next release only if the brand's recent coverage survives into the training window, which argues for durable, well-linked pages over social spikes. The reference-source gap is an earned-media project with a specific target list. Three distinct workstreams, visible only because the tool refused to average them into one number.

The methodology matters more than the score. A credible check infers your category from your homepage, generates the buyer-discovery questions a real prospect would ask, and runs them through the model fresh, recording whether your brand is named and where. Two details separate rigorous checks from theater: the prompts must be buyer-shaped rather than "tell me about brand X" (which any model answers politely and proves nothing), and repeated runs must be expected to vary, which is why a tracker's averaged mention rate outranks any single check. Meev's free checkers publish exactly this method on each checker page, so you know what the number means before you act on it.

The levers differ by mechanism. For training-memory models, the work is earning durable coverage on the sources models learn from: independent publishers, well-maintained reference pages, and consistent entity information across the web; results arrive at the next model release. For retrieval-grounded models, the work is classic answer-engine optimization: rank for the underlying queries, structure pages for extraction, and publish original information worth citing; results can arrive within days. A per-model tool tells you which lever each surface needs, which is the entire point of measuring them separately.