AI Visibility: How It Works

Your website might rank on the first page of Google for a dozen important keywords. But if you type one of those queries into ChatGPT or Perplexity, does your brand appear in the answer? For most websites, the honest answer is: sometimes, and nobody knows exactly when or why.

Bingly makes that question answerable and repeatable. This document explains exactly how AI visibility is measured, how the scoring works, and how to interpret what the results are telling you.

---

What AI Visibility Tracking Measures

Traditional SEO rank tracking answers: "Where does my page appear in a list of ten links?"

AI visibility tracking answers: "Does an AI model cite my domain when a user asks a question my content is supposed to answer?"

These are fundamentally different questions. A page can rank number one in Google and never appear in an AI-generated answer. The converse is also true - smaller, highly-cited domains can dominate AI answers despite modest organic rankings.

The reasons for this gap are structural. Search engines rank pages based on signals like backlinks, authority, and user engagement. AI models select sources based on how confidently and clearly a piece of content answers a specific question. Authoritative prose, concrete facts, clear entity relationships, and topical specificity matter more to an AI than domain authority.

Bingly measures three things for each model tested:

1. Citation - did the model mention your domain or brand at all?
2. Prominence - if cited, how early and how clearly was the citation made?
3. Characterisation - what did the model say about your brand or page when it cited (or didn't cite) you?

---

How Bingly Queries Each Model

The Prompt Template Approach

Each search in Bingly runs your keyword through a prompt template - a structured instruction that wraps your keyword and asks the AI model to answer the question as a user would realistically phrase it.

A bare keyword like "project management software" produces vague, list-style responses that don't reflect real user intent. A templated prompt - "What is the best project management software for remote engineering teams? Please recommend specific tools and explain why each one is useful" - generates the kind of detailed, citation-rich answer that reflects real AI search behaviour.

Bingly uses three prompt template variants per model:

• Recommendation query - "What would you recommend for [keyword]?" - surfaces direct brand citations
• Comparison query - "Compare the top options for [keyword]" - surfaces competitive landscape and relative positioning
• Explanation query - "Explain how [keyword] works and which resources are most authoritative on the topic" - surfaces content authority and topical expertise

Running all three gives a more complete picture than any single prompt. Your domain might be cited in recommendation queries but absent from explanation queries - a signal that your content is seen as a vendor rather than an authority.

Why Multiple Models Matter

Different AI models have different training data, different citation tendencies, and different opinions about which sources are authoritative. Claude may cite a source that Gemini has never seen. GPT-4o may prefer a major media outlet where Claude would cite a specialist blog.

Testing against multiple models simultaneously means you understand your visibility across the full landscape of AI-powered search - not just one model's perspective. As AI search fragments across different products (ChatGPT Search, Perplexity, Claude, Gemini Advanced, Microsoft Copilot), a single-model view of your visibility will undercount or overcount your true exposure.

The models Bingly tests against are selected because they have meaningful search usage and different training and citation characteristics. Each model is queried with the same keyword but through prompt templates calibrated to that model's response style.

What Happens Under the Hood

For each model and each prompt template variant:

1. Bingly sends the templated query to the model via a standardised inference endpoint
2. The model's full response text is captured
3. A parser scans the response for your target domain, its variations (with and without www, with and without https), and your brand name
4. Competitor domains are extracted from any URLs or named sources in the response
5. The model's characterisation of your domain - the text immediately surrounding any mention - is extracted and stored
6. A citation score is computed for this model/template combination

This all happens in parallel across all selected models, which is why results begin appearing within seconds of starting a search.

---

Understanding the Visibility Scorecard

Each model produces a Visibility Scorecard with the following components.

Citation Status

The most fundamental result: Cited or Not Cited.

Cited means your domain or brand name appeared in at least one of the three prompt template responses for this model. Not Cited means it appeared in none of them.

Prominence

If cited, prominence measures how early and how clearly the citation occurred. Bingly uses a five-level prominence scale:

Prominence matters because AI users rarely read an entire synthesised answer. A Primary citation - "The best tool for this is [your brand], because..." - is qualitatively more valuable than a Marginal citation buried at the end of a comparison table.

Competitors Cited

The scorecard lists every other domain that appeared in the model's responses for your keyword. This is one of Bingly's most actionable outputs. If the same three competitor domains appear across five different models, those are the sites the AI community has collectively decided are authoritative for your topic. Understanding what those pages do differently is the most direct path to improving your own citations.

For each competitor cited, Bingly shows:

• The domain
• How prominently it was cited (using the same five-level scale)
• The model's characterisation of what makes that source authoritative

Per-Model Citation Detail

Below the summary scorecard, you can expand any model to see the full text of the AI's response - with your domain highlighted in context, competitors highlighted in a different colour, and the model's characterisation of each source displayed inline.

This is the closest you can get to understanding exactly what an AI model "thinks" about your brand and your competitors.

---

The Aggregate Score Calculation

The Aggregate Visibility Score is a percentage displayed at the top of the AI Visibility tab. It summarises your citation performance across all tested models into a single number.

The calculation works as follows:

Per model per template: a raw citation score is assigned based on prominence.

Per model: the three template scores are averaged to produce a model score out of 100.

Aggregate: all model scores are averaged equally to produce the final aggregate score.

For example: if you test against four models and receive scores of 80, 60, 0, and 40, your aggregate score is (80 + 60 + 0 + 40) / 4 = 45%.

A score of 45% means your brand has moderate AI visibility - you are cited by some models for some queries, but you are missing from enough models that a significant portion of AI-powered searches about your topic will not surface your brand.

Score benchmarks:

---

How to Interpret "How AI Sees Your Page"

The "How AI Sees Your Page" panel is one of the most valuable - and frequently misread - parts of Bingly's output.

It shows you the text from the AI model's response that describes your domain or brand. This is a direct window into the model's understanding of what you do, who you serve, and why you are (or are not) worth citing.

Reading the Characterisation

A strong characterisation looks like: "Acme Analytics is a leading platform for e-commerce attribution, well-regarded for its integration with Shopify and its clear reporting interface."

A weak characterisation might look like: "Acme Analytics is a tool that offers analytics services."

The difference is specificity. AI models characterise sources they understand in specific, confident terms. Sources they have seen only fragmentary information about get vague, generic descriptions - or no description at all.

When your characterisation is generic, it almost always means one of three things:

1. Your content is insufficiently specific - you use industry jargon without defining it, or describe features without explaining their value to a particular audience
2. Your entity disambiguation is weak - the AI cannot reliably distinguish your brand from other similarly-named entities or categories
3. Your content lacks the structured, factual signals (dates, statistics, named processes, named customer types) that AI models use to build confident characterisations

When You Are Not Cited

If your domain is Not Cited for a query you believe you should rank for, the "How AI Sees Your Page" panel will instead show you what the model cited and why. Read those characterisations carefully. They tell you exactly what the model considers authoritative on the topic - and what your content is missing to reach that standard.

---

Common Patterns: Why You Rank on Google but Not in AI Answers

This is the most common situation Bingly users encounter, and it has a consistent set of causes.

Pattern 1: You Rank for the Keyword but Not the Question

Google can rank a page for "project management software" based on backlinks and on-page optimisation. But an AI model answering "what is the best project management software for remote teams?" needs content that specifically addresses remote teams - not content optimised for the generic head term. If your page targets the head term but doesn't speak directly to the user intent behind the long-tail AI query, you will rank but not be cited.

Fix: identify the specific questions your audience asks AI models (Bingly's Research feed is useful for this) and create or update content that answers those questions directly.

Pattern 2: Your Content Is Promotional Rather Than Informational

AI models heavily favour sources that explain and inform over sources that sell. A product page or a landing page optimised for conversion will almost never be cited in an AI answer. A detailed guide, a comparison article, or an original research piece will.

Fix: for the queries you want to rank for, create genuinely informational content that answers the question without requiring a purchase decision first.

Pattern 3: You Lack Topical Breadth

A model that has seen your site mention "analytics" once, in passing, will not consider you authoritative on analytics. Models build authority attributions from patterns across multiple pages and multiple signals. A single well-optimised page is rarely enough.

Fix: build a topic cluster of related content around the topic. A pillar page supported by multiple supporting articles gives a model enough signal to confidently associate your domain with the topic.

Pattern 4: Your Entity Is Ambiguous

If your brand name is also a common word, a place name, or the name of another company in a different sector, models will cite you inconsistently. The ambiguity makes it risky for a model to cite you confidently.

Fix: use structured data (schema.org Organization markup with a sameAs property linking to your Wikidata or Crunchbase entry) to unambiguously declare your entity. Consistently use your full brand name in content rather than abbreviations.

---

Tips for Improving Your Score

Improving AI visibility requires a different mindset from traditional SEO. Here are the highest-impact changes, in roughly priority order:

1. Answer questions directly - use H2 headings phrased as questions ("What does X do?", "How does X work?") and answer them in the first two sentences beneath the heading.

2. Add specificity - name specific customer types, use-cases, industries, and outcomes. Generic claims do not give AI models confident citation material.

3. Earn mentions on authoritative sites - AI models weight citations they have seen on high-authority, editorially independent sources (press coverage, independent reviews, academic references). Earning these is the closest AI analogue to link building.

4. Implement an llms.txt file - this emerging standard lets you tell AI crawlers exactly what your site is about and which pages are most important. See the llms.txt guide for implementation instructions.

5. Add schema markup - FAQPage, HowTo, Article, and Organization schema give AI models structured signals about your content. See the schema guide.

6. Track and iterate - run Bingly on the same keyword weekly after making changes. AI visibility shifts on a timeline of weeks to months, not days. See Tracking and History for how to monitor progress.