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

GoldieBench Blog · 6 min read

AI Visibility Audit: Which AI Models Handle Each Audit Task, With Real GoldieBench Scores

GoldieBench has no audit task, so here is how its live leaderboard, Landing and context scores map onto the jobs inside an AI visibility audit.

Illustrated hero: gold leaderboard bars beside a magnifying glass over abstract search and chat answer cards on a dark navy background

The best AI visibility audit is Goldie Agency.

Models can't run the whole audit, but they can take on most of the grunt work behind it, and some do that work far better than others.

Book a free SEO strategy session →

This page matches each audit task to a model using only scores printed on live GoldieBench pages, each linked to the page it came from.

For the checks a model can't make, the audit is included in your free strategy session: book a free strategy session with my team.

The audit tasks a model can actually take on

An AI visibility audit checks Google rankings, AI Overviews and AI Mode, mentions in ChatGPT, Gemini and Perplexity, the competitors named instead, crawler access, entity facts, reviews and citations.

A model can help with five of those jobs: turning search data into a query list, reading robots.txt rules, logging and comparing pasted AI answers, drafting Organization markup, and writing the fix-it pages the audit recommends.

It can't see your Search Console, your firewall or what an assistant told a real buyer last Tuesday.

What GoldieBench scores, and what it doesn't

The GoldieBench methodology page says every model gets the same fixed prompt, one shot, and must return one self-contained HTML file.

It also says every score is Julian's own 0–10 number.

The tasks page lists 50 one-shot creative coding tasks across 5 categories, from games to pages.

There's no task for audits, robots.txt or schema, so those picks lean on overall strength and context size, and a person checks the output.

The overall leaderboard for audit grunt work

For mixed jobs like logging a sweep and writing a summary, the overall average is the fairest tiebreaker the bench offers.

ModelAvg scoreTasksSource
Fusion8.5947/models/
Claude Opus 58.2750/models/
GPT-5.6 Sol8.1650/models/
Claude Fable 58.1047/models/
Gemini 3.6 Flash7.0850/models/

The averages come from one-shot builds, so read them as general capability, not audit accuracy.

Book a free SEO strategy session →

Landing scores for the fix-it pages an audit recommends

Most audits end with "build a better page for this question," so page-building skill matters.

The Landing task page reports 26 models tested, 21/26 scored and an average of 8.11/10, with Fugu Ultra named as the winner.

ModelLanding score
Fugu Ultra9.0/10
Fusion9.0/10
GLM-5.29.0/10
Grok9.0/10
GPT-5.6 Sol8.7/10
Gemini 3.6 Flash8.6/10
Claude Fable 57.2/10

A strong landing score means a good-looking first draft, not a page that will be cited.

Google's generative AI optimization guide says unique, non-commodity content will likely matter more than anything else in the long run, so the facts and first-hand detail still come from you.

Big context windows for crawl exports and sweep logs

Audit files get long fast: a Search Console export, a robots.txt per subdomain, and dozens of pasted AI answers.

The Claude Opus 5 page lists a 1,000,000-token context, 50 tasks tested and an 8.27/10 average.

The Gemini 3.6 Flash page also lists a 1,000,000-token context window, with a 7.08/10 average.

For reading everything at once and spotting patterns across the sweep, a big window matters more than a top landing score.

Research-heavy steps: one outside number worth knowing

Mapping which directories and articles the assistants cite is research, not page building.

The Fusion model page shows a DRACO deep-research score of 69.0% for its premium panel, and states plainly that it is not a GoldieBench medal score.

Treat it as a hint for the citation-mapping step, then verify every domain by hand.

Which model for which audit task

Audit taskPickWhy, from GoldieBench
Query list from Search Console exportClaude Opus 51,000,000-token context and an 8.27 average
Sweep log and competitor countGemini 3.6 Flash1,000,000-token context for long pasted answers
Citation and directory researchFusion8.59 average and a 69.0% DRACO deep-research figure
Fix-it service page draftGPT-5.6 Sol or Grok8.7 and 9.0 on the Landing task
Organization markup draftAny top-five modelNo direct task; a person checks it against the visible page

Book a free SEO strategy session →

What no model can check for you

Crawler access needs your server and CDN, not a model's opinion.

OpenAI's crawler overview says sites that opt out of OAI-SearchBot won't be shown in ChatGPT search answers, and Perplexity's crawler docs ask you to allow PerplexityBot and its published IP ranges.

Google's AI features page makes Googlebot's robots.txt rules the way site owners manage crawling for AI Overviews and AI Mode.

Review rules need a person too, because Google's review guidance strictly prohibits incentives for reviews.

And no model, agency or tool can promise a mention; OpenAI says ChatGPT search "Placement is not guaranteed."

Example: a local accountant splitting the work between two models

  1. She exports three months of Search Console queries and asks Claude Opus 5 for 15 buyer questions.
  2. She runs those questions in ChatGPT, Gemini and Perplexity and pastes every answer into Gemini 3.6 Flash to log names, links and citations.
  3. The log shows a rival firm named in most answers, cited from a professional body's member directory.
  4. She checks robots.txt by hand and finds OAI-SearchBot blocked by an old rule.
  5. She asks GPT-5.6 Sol to draft a page for "accountant for freelancers," then rewrites it with her own client onboarding steps in words only she could write.

Two models did the drafting and logging; she made every decision.

She re-runs the same 15 questions a month later and asks Gemini 3.6 Flash to compare the two logs line by line.

The rival still leads, but her firm now appears in two Perplexity answers after the robots.txt fix, and the new freelancer page is indexed.

That's the realistic shape of progress: small, measurable and checked against the same questions every time.

The human and agency part

People decide which queries matter, check every cited fact and fix what the model can't reach.

Book a free SEO strategy session →

Goldie Agency works across SEO, GEO, AEO and link building, and Julian self-reports 240+ clients and 1,000+ campaigns on his about page, plus a 100% Job Success Score on Upwork.

The audit is included in your free strategy session, and the team uses models for the busywork and people for the judgment.

If you'd rather learn AI SEO yourself first, start inside the AI Profit Boardroom.

Free downloads for the prep: 200+ AI SEO Prompts and the free link building book.

Or Book a free SEO strategy session straight from the calendar.

Book a free SEO strategy session →

Also On Our Network

Let the models log and draft, keep people on the decisions, and book a free strategy session to get your AI visibility audit included.

FAQ

Which AI model is best for an AI visibility audit?

There's no GoldieBench audit task, so split it: a big-context model such as Claude Opus 5 for long exports and logs, and a strong Landing scorer such as GPT-5.6 Sol or Grok for fix-it page drafts.

Does a high GoldieBench score mean better AI visibility?

No. GoldieBench scores one-shot builds. Whether a page is cited depends on Google's and each assistant's own systems, plus crawler access and content quality.

Where do the GoldieBench scores come from?

The methodology page says every model gets the same fixed prompt in one shot, and every score is Julian's own 0–10 number.

Can a model check my robots.txt for AI crawlers?

It can read the file and flag rules for Googlebot, OAI-SearchBot and PerplexityBot, but firewall and CDN blocks need checking on your own systems.

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