AEO Agency UK Work And Which AI Models To Use For It, With Real GoldieBench Scores For Page Drafting

An AEO agency UK clients hire still writes briefs, drafts schema and builds landing pages, and the AI model doing that drafting changes how much editing the humans do afterwards.

This page matches those tasks to models using only scores published on live GoldieBench pages, quoted with the page each came from.

For the human half of the job, book a free SEO strategy session with my team.

What GoldieBench measures, and what it does not

The GoldieBench models page describes the method: the same fixed prompts, one shot each, scored 0–10 by Julian.

The leaderboard lists 50 tasks and 33 models, split into Games, Visuals, Sims and Pages.

The Pages category has 3 tasks, and it is the closest match to SEO work because it tests whole marketing pages.

There is no GoldieBench task for schema markup, content briefs or outreach research.

So treat the scores below as a guide to page-building quality, not a verdict on every AEO job.

The AEO tasks a model actually touches

Landing-page scores: the closest proxy for AEO answer pages

The GoldieBench Landing task page asks every model for a modern marketing landing page in one shot.

That page reports 26 models tested, 21 of 26 scored, an average of 8.11/10, and Fugu Ultra as the winner.

A landing page that looks premium is only half an answer page.

The page still needs your facts, your proof and a direct answer at the top, which no benchmark scores.

Overall leaderboard: which models hold up across all tasks

On the GoldieBench leaderboard, ranked by average score across all tasks, the top of the table reads:

GPT-5.6 Sol is the highest-ranked model whose strongest category is Pages, which makes it a sensible default for page drafts.

Matching models to AEO jobs

• Answer and service pages: start with a Pages-strong model such as GPT-5.6 Sol (8.7/10 on Landing) or a 9.0 scorer such as GLM-5.2 or Grok, then edit for accuracy.
• Schema drafts: no GoldieBench score exists, so use any capable model and validate every value against the visible page, as Google's Organization structured data guide expects markup to reflect the page.
• Content briefs: no GoldieBench score exists either; choose the model your team already trusts and check it never invents facts.
• Audits: the model matters less than the agent's ability to fetch robots.txt and headers reliably.

What the models cannot do for AEO

A model cannot open your robots.txt to the right crawlers if a CDN still blocks them.

OpenAI's crawler overview says sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, and its ChatGPT search article adds that placement is not guaranteed.

A model also cannot manufacture authority.

Google's generative AI optimisation guide says chasing inauthentic mentions isn't as helpful as it seems, and it puts non-commodity, first-hand content above every other suggestion.

That first-hand experience comes from the business, not the model.

A small-business workflow using these scores

A Nottingham garden landscaper:

1. Lists 15 customer questions, such as whether a patio needs planning permission.
2. Uses a Pages-strong model to draft five answer pages from the owner's voice notes.
3. Edits each draft for accuracy and adds real project photos.
4. Drafts LocalBusiness markup with any model and checks every field against the contact page.
5. Checks the Bing AI Performance report a month later for citations.

A York wedding photographer:

1. Asks the agent to confirm PerplexityBot and Claude-SearchBot are allowed.
2. Drafts a venue guide page with a high Landing scorer, then rewrites it with first-hand notes from real shoots.
3. Pitches two regional wedding blogs personally for genuine features.

Run your own fair test before committing to a model

GoldieBench uses one fixed prompt per task, so copy that discipline for your own pages.

1. Write one brief for a real service page, with your facts and three customer questions.
2. Give the identical brief to two or three models from the tables above.
3. Score each draft from 0 to 10 on accuracy, clarity of the opening answer, and how much editing it needs.
4. Check whether any model invented a fact, a price or a claim you can't support, and mark it down hard.
5. Keep the winner as your default and rerun the test when a new model tops the leaderboard.

Accuracy beats polish for AEO, because AI answers pull facts, not design.

Google's AI features documentation also reminds site owners that structured data should match the visible text, so a model that invents details in markup fails the test.

The human part: where the agency earns its place

Benchmarks pick the drafting engine.

People still decide which questions deserve a page, which crawlers to allow, which outlets to approach, and how to report results accurately.

Google's advice on hiring an SEO says no one can guarantee a #1 ranking, so judge any provider on process and evidence.

Goldie Agency works with UK businesses remotely as an SEO, GEO, AEO and link building agency, and describes its outreach as manual outreach, no PBNs.

By Julian Goldie's own published numbers, that work spans 1,000+ SEO and link building campaigns, and he shares AI search tests with 425K+ YouTube subscribers.

If you would rather learn AI SEO yourself, the AI Profit Boardroom teaches the same model-plus-agent workflow; for the GEO side, see our GEO agency UK model guide.

Also On Our Network

• Agent Operatorsthe seven-step operator process with robots.txt lines, schema types and reporting
• AI Income Deskwhether hiring for answer engine optimisation pays off for a small UK business
• AI Tool Verdictfive ways to get answer engine optimisation done, each rated out of ten
• agentos.guidefive copy-paste agent prompts that do the answer engine groundwork first

Pick a Pages-strong model for drafts, keep people on facts and outreach, and book a free strategy session with the AEO agency UK businesses use for the rest.