⭐ Get the Agent OS + join 3,400+ founders inside the AI Profit Boardroom → Join AIPB

GoldieBench Blog · 6 min read

AI Visibility Tracking Engine by Engine: What to Log for ChatGPT, Gemini, Perplexity, Claude and Google AI, Plus the Model to Analyse It

Why one blended AI score misleads, how to log each engine separately, and which Goldie Bench model to use when you read the results.

Illustrated hero: five separate gold columns of answer cards beside a small leaderboard on a dark navy background

AI visibility tracking measures how often each AI engine names your brand, cites your pages and frames you well, and the key word is each: ChatGPT, Gemini, Perplexity, Claude, Copilot and Google's two AI features all build answers differently.

Book a free SEO strategy session →

This page explains why one combined score misleads, what to log per engine, how to set it up, and, near the end, which AI model to use when you analyse the log.

Want the setup done for you? Book a free SEO strategy session with Goldie Agency.

AI visibility tracking, defined

You choose a fixed set of buyer questions, put them to each engine on a schedule, and record the result in a structured log.

The log captures six signals: mention, citation, share of voice, sentiment, branded versus unbranded prompt, and engine.

Rank tracking assumes a stable list of positions. AI answers don't have one, so presence across many samples replaces position on one page.

Why every engine needs its own column

Engines differ in where they fetch pages and how they write replies. A brand can lead on one and be missing on another.

  • Google AI Overviews and AI Mode: Google's AI features page says the two may use different models and techniques, so their responses and links vary. Treat them as two engines.
  • ChatGPT search: OpenAI names OAI-SearchBot as the crawler behind ChatGPT search results; opt it out and your pages drop from those answers.
  • Perplexity: Perplexity says PerplexityBot surfaces and links sites in its results.
  • Claude: Anthropic says Claude-SearchBot indexes content for search, and disabling it may cut visibility.
  • Copilot and Bing: Bing publishes its own citation data, covered below.

Average them into one figure and a strong ChatGPT month can hide a collapse in AI Mode.

What to log per engine

SignalHow to record itWhat a change usually means
Mention1 or 0 per answerBrand awareness on that engine is moving
CitationThe cited URL, or blankThat page is or isn't trusted as a source
Share of voiceYour mentions over all brands namedCompetitors are gaining or losing ground
SentimentPositive, neutral or negative plus a noteSource pages describing you have changed
Prompt typeBranded or unbrandedReputation versus discovery

Book a free SEO strategy session →

Report every rate with its sample size, per engine, per prompt type.

Branded and unbranded, engine by engine

Split every engine's column again by prompt type. Branded prompts tell you how an engine describes you once a buyer already knows your name. Unbranded prompts tell you whether that engine introduces you to buyers who don't.

The two often disagree. An engine can describe you warmly when asked about you by name and still leave you out of every category answer.

When budget or time is short, prioritise the unbranded split. That's where new customers come from.

Setting up a per-engine tracker

  1. Write 24 prompts: 16 unbranded, 8 branded. Use your customers' wording.
  2. List three to five real competitors.
  3. Pick five engines your buyers use and give each its own column.
  4. Run every prompt on every engine in one day, twice if you can.
  5. Repeat monthly with the same prompts and compare.

Twenty-four prompts across five engines is 120 rows a sweep: manageable by hand and big enough to show movement.

Treat answers as samples, not facts

According to Google, its AI features can split one question into many related searches across subtopics, a method it calls query fan-out. Small differences in that step change which pages get pulled in.

Bing adds that citation activity can shift with evolving AI models, competing content, freshness and demand.

So act on a change only when it repeats across two sweeps. One odd answer is noise.

What Google and Bing report directly

Two engines publish first-party numbers about your own site:

  • Google: the Generative AI performance report in Search Console logs each time a link to your site appears in AI Overviews or AI Mode, grouped by URL, country, device and date. Its help page lists no query or click columns.
  • Bing: AI Performance shows citations across Copilot and Bing AI summaries, sampled grounding queries and, since June 2026, Citation Share per query.

ChatGPT, Perplexity and Claude don't publish an equivalent report for site owners, which is why the prompt log carries those columns.

Book a free SEO strategy session →

Tools that automate per-engine sampling

Tools differ mainly in which engines they cover, going by their own sites:

  • Ahrefs Brand Radar lists Google's AI Overviews and AI Mode alongside ChatGPT, Perplexity, Gemini and Copilot.
  • Semrush's AI Visibility Toolkit focuses its prompt tracking on AI Mode and ChatGPT.
  • Peec AI lets you segment by model, so per-engine columns come built in.
  • Otterly.AI reaches seven engines, Claude and Copilot among them.

For the ranked list, see the best AI visibility tools. And keep Google's generative AI guide in mind: tools outside Google don't get into its ranking or AI systems.

Which model to use when you analyse the log

You can't choose which model answers your buyers. You can choose which one reads your 120-row log and writes the monthly summary.

On the live Goldie Bench leaderboard, read on 16 November 2026, Fusion leads overall at 8.59, Claude Opus 5 is second at 8.27 and GPT-5.6 Sol scores 8.16 and is marked best at Pages.

Those are Julian's 0 to 10 scores averaged across build tasks, per the methodology. They rate building, not search visibility, so use them to pick a strong general analyst, then check a sample of its counts by hand.

A worked example: a bakery in Bath

A bakery tracks 24 prompts such as "best sourdough in Bath" and "birthday cakes delivered in Bath" across five engines.

Month one: named on Perplexity and ChatGPT, absent from AI Mode, and AI Overviews didn't trigger for most prompts.

The log shows AI Mode citing a local food blog that lists three competitors. The fix is earning a genuine feature on that kind of site, not rebuilding the bakery's own pages.

Month two: same prompts, same day of the month. If AI Mode still skips the bakery but Search Console shows new AI impressions for its cake page, the prompts may be phrased differently from how locals really search, so the owner adds two variants in a separate tab.

If a pattern points to a wider problem, an AI visibility audit checks access, entity facts and citations end to end.

My verdict: one log, one column per engine

Track each engine separately, pair the log with the Google and Bing reports, and judge trends across two or more sweeps.

To learn the full method, join the AI Profit Boardroom. To have it done, book a free SEO strategy session. Members who want to compare models on their own logs can also do that inside the Boardroom.

Book a free SEO strategy session →

Also On Our Network

Give every engine its own column, compare like with like each month, and get a free strategy session if you'd like the team to read your AI visibility tracking with you.

FAQ

Why track each AI engine separately?

Each engine fetches and writes differently. Google says even AI Overviews and AI Mode may use different models, so one blended score hides real changes.

Which engines should a small business track?

The ones its buyers use. Most start with ChatGPT, Gemini, Perplexity and Google's AI Overviews and AI Mode.

Does any AI engine publish citation data for my site?

Google's Search Console shows AI impressions, and Bing Webmaster Tools shows citations across Copilot and Bing AI summaries. No equivalent report is documented for ChatGPT, Perplexity or Claude.

Do Goldie Bench scores show which AI cites me most?

No. Goldie Bench scores models on build tasks. Use it to pick a model for analysing your log, not to predict citations.

How many prompts are enough per engine?

Around 24 prompts per engine gives enough rows to see trends while staying manageable by hand.

What if an engine never cites me?

Check crawler access first. OpenAI, Perplexity and Anthropic each document a search crawler that must be allowed in robots.txt.

The same stack Julian uses

Run this stack yourself.

Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 3,400+ founders shipping with it every day all live inside the AI Profit Boardroom.

3,400+founders
258documented wins
38countries
4coaching calls a week