How Lume scores AI visibility — and why the number has to be deterministic

Most business owners who come to Lume have the same question. They want to know: when someone asks ChatGPT or Google's AI about my type of business in my area, do I show up?

It's a reasonable question. The answer turns out to be harder to compute than it sounds, and the way you compute it matters a lot.

Why the score has to be a fact

When we built the first version of Lume's scoring, we made a mistake. We let a language model produce the headline number — the overall AI visibility score — by asking it to synthesize everything it knew about a brand across a set of prompts. The result felt intelligent. It also varied by a few points every time you ran the same audit. Which version was right?

Neither, really.

This is the same problem any scoring system runs into when it lets a probabilistic model produce a deterministic-looking output. Ahrefs Domain Rating is computed by a link-graph algorithm. Semrush Site Health is computed by a weighted-issue-count formula. Google's Lighthouse is a weighted average of six metrics. In each case, the number comes from rules, not from an AI's read of the situation. The AI handles the narrative — the "here's what this means for your business" explanation — but it never touches the number itself. That's why you can run the same audit twice and trust the result.

We rebuilt Lume's scoring around the same principle.

What we measure

AI visibility for a small business comes down to four questions.

1. Are you being cited?

When someone asks an AI assistant a question your business should answer, does your brand appear in the response? Citation rate — how often your brand appears across a set of relevant prompts — is the most direct measure of AI visibility. We run prompts across ChatGPT, Gemini, Claude, and Google's AI Mode. We count the responses where your brand shows up.

2. What's your share of voice?

Appearing in one out of ten relevant responses is different from appearing in eight out of ten. Share of voice is the ratio: how often does your brand appear compared to the total space available across the prompts we test? A local plumber with a high citation rate and high share of voice is the brand AI assistants have converged on as the credible answer. That's the goal.

3. Which platforms cover you?

Not all AI platforms see your brand the same way. A brand that appears consistently in ChatGPT responses but is invisible to Gemini has a platform coverage gap. We track citation rate and share of voice per platform — ChatGPT, Gemini, Claude, and Google AI — because the training data, grounding sources, and retrieval patterns differ across them.

4. Where do you appear?

Citation position matters. Being named as the primary answer is different from appearing as a footnote in a longer response. Position score captures where in an AI response your brand typically surfaces.

Each of these four sub-scores — citation rate, share of voice, platform coverage, position — is computed from crawl data and prompt-run results, not from an AI's judgment call. We combine them using a weighted formula. You can run the same audit on Monday and Tuesday and explain exactly why the number changed, because the inputs are the same.

The confidence layer

Here's where most scoring tools stop. They run the prompts, count the citations, produce a number.

The problem is that not everything is measurable on every audit run. A site protected by Cloudflare challenges can block our crawler. A brand with no social presence has zero data for certain sub-scores. If we scored those brands the same way we score a fully-accessible one, we'd produce false-low numbers — and a business owner would look at a low AEO score and think the problem is their brand, when the problem is actually incomplete data.

We added a confidence layer that handles this explicitly. When we can't measure a sub-component, we exclude it from the weighted average and renormalize across what we could measure. We surface a confidence indicator — "we measured 3 of 4 pillars; site crawl blocked" — so the score you see is an honest summary of what was measurable, not a verdict on what we couldn't access.

A score that says "your citation rate is 4 out of 5, but we couldn't access your site" is more useful than a score that says "you scored 48." Honesty about measurement coverage is part of getting the score right.

Why measurement alone isn't enough

In May 2026, Google published its first official AI optimization guide — and its framing surprised a lot of people. Google said that AEO (Answer Engine Optimization) and GEO (Generative Engine Optimization) are "still SEO." Not a new category. Not a separate discipline. The same core index and quality signals that earn rankings are what earn citations in AI answers. Source: Google Search Central — covered plainly by Search Engine Journal. That matters for how you think about measurement: if AI visibility is an extension of SEO, not a replacement for it, the fundamentals still drive the score.

At the same time, there are now several tools that measure AI visibility well. Enterprise-tier platforms with per-platform dashboards, citation-source classification, competitor tracking. Some are priced at $399/month; some at $99/month.

What they share: they measure and stop. You get a score, a trend line, a breakdown by platform. The question of what to do about a low score is still yours.

Knowing your citation rate is 12% is useful. Knowing that your citation rate is 12% because your schema markup is incomplete, your Google Business Profile information is inconsistent across directories, and you have no content that matches the way AI prompts retrieve local service information — that's actionable.

Lume measures all four pillars, then generates fixes: schema improvements, content gap audits, business profile corrections, social presence recommendations. The measurement is the first step. The fixes are the point.

AI-referred traffic converts at 14.2% versus 2.8% for Google organic — roughly five times higher. Source: Emarketed, 2026 ChatGPT referrals to brand homepages roughly doubled after May 7, 2026, when ChatGPT started including branded links in responses. Source: Profound The traffic from AI search is real, it's growing, and it converts better than what most small businesses are already optimizing for.

The gap is between knowing you have a problem and knowing what to do about it. The score is how you find out where you stand. The fixes are how you close the gap.

What good looks like

A business with strong AI visibility across all four pillars tends to have a few things in common:

• Consistent, accurate information about what they do and where they operate — across their website, their Google Business Profile, and third-party directories. AI systems ground their responses in structured data. Inconsistency creates ambiguity, and ambiguity gets you left out.
• Original content that answers questions the way people actually ask them. Not keyword-stuffed service pages — content with a point of view, specific to the problems they solve. AI systems prioritize "unique data, original insights, fresh perspectives over repetition." Source: Green Flag Digital
• Schema markup. Structured data tells AI systems exactly what a business is, what it offers, and what it costs. Schema markup increases source citations by 30%. Source: Exposure Ninja, 2026
• Citations from multiple sources. AI engines build consensus before citing a business as factual. A business that appears in multiple credible directories, review platforms, and third-party mentions is more likely to be cited than one that exists only on its own website.

None of this is mysterious. It's the same quality signals that have always determined trustworthiness — applied to the way AI systems retrieve and present information.

The methodology page

If you want to see the full scoring formula — the exact sub-components, the weights, the confidence calculation — it's at getlumeai.com/methodology.

The goal of that page is the same as the goal of this post: if you ran the same audit twice on the same day, you should be able to explain why the numbers are the same or why they changed. Scores you can't explain aren't scores. They're just numbers.