How Foreground works

What is AI search visibility?

More people now ask ChatGPT, Gemini, and similar tools questions like "what's the best accountant near me" or "who does custom furniture in Rotterdam" instead of typing them into Google. Those AI answers don't return a list of ten blue links you can scan for your own name — they return a short, confident paragraph that either mentions your business, mentions a competitor instead, or mentions neither.

AI search visibility is whether your business shows up in those AI-generated answers. It's the same underlying idea as traditional SEO (does Google rank you?), applied to a different surface — one where there's no results page to check by hand, because the "result" is a sentence a model generated on the fly. Think of every answer as having a foreground with room for only a few names — this is how you find out whether yours is one of them.

How a scan works

When you run a scan, Foreground sends 5 different prompts — phrased the way a real customer would actually ask, not just "tell me about {your brand}" — to 4 AI models (OpenAI's GPT-5 mini, Google's Gemini 3 Flash, Anthropic's Claude Haiku 4.5, and xAI's Grok 4.6), for 20 checks total per scan. Some prompts ask for a recommendation directly ("what's the best X for Y?"), some ask for a comparison against your named competitors, some ask "who are the leaders in X category?" — covering the range of ways someone genuinely deciding between options might ask.

Every answer is grounded in a live web search rather than only the model's static training data — each model is queried directly against its own provider's API (OpenAI, Google, Anthropic, xAI), using that provider's own web-search tool. See "Limitations, honestly stated" below for a couple of real coverage gaps that are worth knowing about.

For each of the 20 responses, Foreground checks whether your brand name (or a domain-derived alias) is mentioned, whether it's the first business mentioned, and which competitors — if any — appear instead or alongside it. This is presence detection, and it's intentionally conservative — very short or common-word brand names are flagged as ambiguous rather than guessed at, since a naive text match on a name like "Best" would match almost every response regardless of whether your business was actually meant.

Presence alone doesn't say whether a mention was a genuine recommendation or just a competitor beating you to the punch, so every check where your brand is actually mentioned also gets run through a second pass: a sentiment judge that classifies the mention as recommended, neutral, negative, or comparison-only. This runs automatically on every scan, right after the main 20-check pass completes, and shows up as a summary on the Overview tab plus a per-check badge in the detailed breakdown.

How the 0-100 score is calculated

Across the checks that actually completed (a failed API call isn't counted as "invisible" — it's excluded, since we don't have data either way), each check earns credit based on exactly where your brand appeared relative to competitors in that answer: full credit for being the first business mentioned, decaying partial credit for 2nd, 3rd, or further down — being in the conversation at all is a real signal, just a weaker one than being the top answer. Never being mentioned earns nothing.

Checks aren't all worth the same, either. A prompt phrased like a real buying decision ("what's the best X for Y?") counts for more than a broad "who are the leaders in this category?" prompt, since the former is closer to how a customer actually decides — so ranking first on a direct-intent question moves your score more than ranking first on a general one. The bands you'll see on your dashboard:

• Leading (80-100) — you're the first answer in most checks.
• Visible (50-79) — you show up regularly, sometimes first.
• Weak (1-49) — you're mentioned occasionally, rarely first.
• Invisible (0) — not mentioned in any completed check.

If too few checks complete — a real outage, not a real result — Foreground shows no score at all rather than a misleading one computed from too little data. Your dashboard shows exactly how many of the 20 checks succeeded, so you can judge how much to trust a given score. A 0 always means "checked and genuinely not mentioned," never "we couldn't check."

What to do with a low score

A low score means you're not the one in the foreground when your customers ask AI these questions. A few directions that tend to help:

• Check what these models are actually reading. Since answers are grounded in live web search, models tend to lean on sources that are easy to find and clearly written — review sites, industry directories, comparison articles, and your own site's content. If those sources don't mention you clearly, an AI model has nothing to cite.
• Make your own site answer the question directly. Content that plainly states what you do, who it's for, and how you compare to alternatives is easier for a grounded search to pick up than vague marketing copy.
• Look at who's beating you. Your scan's competitor breakdown shows which competitors get cited instead, and how often — that's a direct signal of who the AI models currently consider the default answer in your category.
• Track it over time, not just once. A single scan is a snapshot; AI models' outputs shift as the web changes. Re-scanning periodically shows whether changes you make are actually moving the needle.
• Use the deeper advice on a scan. Pro accounts can generate a grounded, scan-specific breakdown of what's driving your result and what to focus on first, instead of generic SEO advice — see pricing.

Limitations, honestly stated

AI model outputs aren't fully deterministic or fully within anyone's control, including ours. A score reflects a real check against real model outputs at a real point in time — it's a strong directional signal, not a certified measurement, and it can shift between scans even with nothing changed on your end. Treat it the way you'd treat any single analytics snapshot: useful, worth tracking over time, and not the whole picture on its own.

On the "4 AI models" claim, specifically: GPT-5 mini, Gemini 3 Flash, Claude Haiku 4.5, and Grok 4.6 are each called directly against OpenAI's, Google's, Anthropic's, and xAI's own APIs respectively, each grounded using that provider's own web-search tool — so this measures each model in its own real product, not a third party's approximation of it. There's one asterisk worth naming: Gemini's citations come back as Google's own grounding-redirect links rather than the original source URL, so while Gemini's answer text is checked the same way as the others, its citation links can't be matched against your own site's domain the way the other four providers' citations can. Foreground also doesn't currently check Google's AI Overviews/AI Mode or Microsoft Copilot — AI-answer surfaces this scan doesn't cover. We'd rather state all of this plainly than let "5 models" imply more or less coverage than what's actually being measured.