AI SEO Firm Work: Which AI Models Draft The Pages, Tools And Assets, Scored On Live GoldieBench Pages

Any AI SEO firm worth hiring in 2026 uses AI models for first drafts, and the model it picks changes how much fixing its people have to do afterwards.

This page shows which models score best on the GoldieBench tasks closest to firm deliverables, using only scores read on live goldiebench.com pages on 11 Nov 2026.

It then covers what the human side of a firm must still own, and the questions to ask about its AI workflow before you sign.

For the strategy and the safeguards, book a free SEO strategy session with my team.

What GoldieBench actually scores

The GoldieBench leaderboard shows 1310 live demos across 50 tasks and 33 models.

It says the ranking is Julian's own 0–10 scores averaged across every task a model ran.

Each task is one shot, one HTML file out, no iteration, which is close to how a firm uses a model for a first draft.

The Pages category has three tasks; none of them scores schema, briefs or outreach, so read the numbers as page-building skill only.

Firm deliverables and the nearest bench task

Landing scores: the page-draft signal

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

Four models tie at the top, so a firm has real choice for page drafts.

Web-OS scores: the interactive-tool signal

The Web-OS Desktop task asks for a working desktop with windows, a dock and three apps, which tests whether a model can ship working interface code.

It reports 21 of 26 scored and an average of 7.98/10.

• Fusion: 9.0/10, winner.
• Grok: 9.0/10, winner.
• GPT-5.6 Sol: 8.4/10.
• Claude Fable 5, Gemini 3.6 Flash and Kimi K2.7: 8.0/10 each.
• GLM-5.2: 7.5/10.

Claude Opus 5.5's model page lists it at 8.7 on Web-OS with a bronze medal, its best Pages result, against 7.5 on Landing and 7.2 on AIPB Promo.

AIPB Promo scores: the launch-asset signal

The AIPB Promo task asks for an auto-playing motion-graphics advert in one HTML file.

It reports 20 of 26 scored, an average of 7.48/10, and Qwen 3.8 as winner.

Scores on that page include GPT-5.6 Sol at 8.3/10 with silver, Gemini 3.6 Flash at 7.8/10, Claude Fable 5, Fusion and GLM-5.2 at 7.5/10, and Grok at 7.0/10.

The overall table, and why it matters less here

On the leaderboard, Fusion leads with an 8.59 average, then Claude Opus 5 at 8.27, Hermes MoA at 8.17 and GPT-5.6 Sol at 8.16.

GPT-5.6 Sol and GLM-5.2 both carry the "best at Pages" tag there.

Overall averages include games and simulations, so for firm work the three Pages tasks are the better guide.

Across all three, GPT-5.6 Sol is the steadiest: 8.7, 8.4 and 8.3.

What the platforms say no model can skip

• Google's AI features page: a page must be indexed and eligible for a snippet to appear in AI Overviews, with no extra technical requirements.
• Google's generative AI guide: mass-producing pages for every query variation can count as scaled content abuse, however they're made.
• OpenAI's crawler docs: blocking OAI-SearchBot keeps a site out of ChatGPT search answers.
• Perplexity's crawler page: allow PerplexityBot to show in its results.
• Bing's Webmaster Guidelines: GEO doesn't guarantee citations in AI answers.

A model can write a 9.0 page that nobody's crawler is allowed to read.

A model pairing a firm could run

No single model tops every Pages task, so a sensible firm splits the drafting by job.

1. Service and location pages: one of the four Landing 9.0 models, such as Fusion, Grok, GLM-5.2 or Fugu Ultra.
2. Quote tools and calculators: Fusion or Grok, both 9.0 on Web-OS, or Claude Opus 5.5 at 8.7.
3. Second opinion and edits: GPT-5.6 Sol, which stays above 8 on all three Pages tasks.
4. Schema, briefs and outreach research: any strong model, checked line by line by a person, since no task scores this work.

Ask the firm whether it re-checks its choices when new scores land, because the leaderboard changes as models are added.

A firm that still runs last year's model on every task is leaving quality on the table.

Questions to ask a firm about its AI workflow

1. Which models draft our pages, and who edits them?
2. Who checks every fact, claim and statistic before publishing?
3. Where does the first-hand detail come from: our photos, our jobs, our staff?
4. How do you stop one template becoming hundreds of near-identical pages?
5. Does the model ever get write access to our CMS, or only drafts?
6. Who owns the prompts, drafts and outputs when we part ways?

A firm with good answers treats the model as a junior writer, not a publisher.

Vague answers, or a claim that the AI "does it all", tell you who will be fixing the mistakes: you.

The human and agency part

People still decide the topics, add proof, earn links, hold your accounts to the right roles and take responsibility for what goes live.

Google's hiring page is blunt: you're responsible for what any company you hire does, and no one can guarantee a #1 ranking.

Goldie Agency is Julian Goldie's SEO, GEO, AEO and link building agency, and it describes its outreach as "Manual outreach, no PBNs".

Julian's own published numbers put it at 1,000+ SEO and link building campaigns, 240+ clients and a 100% Upwork Job Success Score, with 425K+ YouTube subscribers on his channel.

Book a free SEO strategy session to see how that team splits work between models and people.

Small-business example: a kitchen fitter's new service pages

1. The firm drafts five service pages with a Landing 9.0 model such as Fusion or Grok.
2. A strategist swaps generic lines for the fitter's real job photos and fitting steps.
3. A developer builds a simple layout-planner tool with a strong Web-OS model, then tests it.
4. The owner approves every claim, and the firm logs each page as it goes live.
5. The firm checks robots.txt so OAI-SearchBot and PerplexityBot can reach the new pages.

Prefer to learn AI SEO yourself first? The AI Profit Boardroom covers the drafting workflow end to end.

Also On Our Network

• Agent Operatorsthe client-side access, contract and crawler setup before an AI SEO firm starts
• AI Income Deskwhether a small business should hire an AI SEO firm, a freelancer or nobody yet
• AI Tool Verdictan AI SEO firm scored against freelancers, in-house hires and DIY
• agentos.guidesix Agent OS prompts that build your evidence pack before hiring an AI SEO firm

Let strong models draft, keep people on facts and links, and book a free strategy session when you want help picking an AI SEO firm.