AI Optimizely

Getting more from every credit: practical ways to run Mark agents efficiently

Jon Seal
Jon Seal Nov 18, 2026, 1:57:29 PM 6 min read

Getting the most from Mark agents · Part 2 of 3

In part 1, we found that what Mark agents cost and what they deliver don't always line up. This post covers the practical levers that bring credit use down and the quality of findings up.

7

practical ways to get more from every credit

1

page per template for structural checks

20–25

priority pages on a monthly hit list

2–3

audits a year for each priority page

We ran seven of Optimizely's ready-made Mark agents against live sites. Some produced sharp, useful findings for modest credit usage. Others repeated each other, or spent credits on work that added little value or wasn't central to the task.

The encouraging part is that most of this is within your control. Credit use isn't fixed by the platform. It depends on how each agent is set up: how tightly its instructions are written, which tools it can call, how much it has to work out for itself, and how hard it's asked to think. The ready-made agents are, rightly, built to work for anyone, so they tend to do more than any one organisation might need. They work well out of the box, but there's real value in tailoring them to your organisation.

Here are seven ways to get more from every credit.

1. Decide what you're trying to learn before you run anything

The agent directory is a useful place to start. But picking a handful of agents that look helpful and running them all is the fastest way to run out of credits.

Start instead with the questions the business needs answering. For many organisations, they look like this:

Risk: Are we exposed to risks (for example accessibility, compliance or brand risks)?

AI search: Can AI search tools find, understand and quote our content?

Trust: Does our content earn trust, and is it accurate and current?

Speed: Are we fast enough that speed isn't costing us customers?

Then identify one agent to own each question. Anything an agent produces that doesn't answer one of those questions is cost without value.

2. Remove the overlaps

Each ready-made agent is designed to stand alone, so each one checks the basics for itself. Run several on the same page and you pay for the same findings more than once.

This is what we saw across the agents we tested:

Check GEO Auditor EEAT Checker Accessibility Alt Text Page Performance
Page speed (Lighthouse and real-user data) ✓       ✓
Structured data (schema) ✓ ✓      
Image descriptions (alt text) ✓   ✓ ✓  
Headings and meta tags ✓ ✓ ✓   ✓
Automated accessibility score     ✓   ✓

Fixing the overlaps is relatively simple: give each check to one agent. For example, if you decide image descriptions belong to accessibility, then tell the others to skip it. Copy the ready-made agent, edit its instructions and add a line such as "don't assess image descriptions". That stops the work and reduces the cost.

3. Audit templates, not pages

Most sites are built from a handful of templates: home, landing pages, product or service pages, articles, forms. Many of the issues agents find don't belong to a page at all. They belong to the template, and they repeat on every page built from it.

Links that went nowhere readable

On one site we tested, several links in a key service section pointed to raw internal CMS addresses rather than readable web addresses, the same pattern in every panel.

Accordions that kept screen readers guessing

On another, accordion buttons didn't tell screen readers whether they were open or closed.

Both are template problems. Fix them once and potentially hundreds of pages improve.

So for structural and accessibility checks, run one representative page per template rather than every page. Then re-run that page after each release that touches the template. You get the same coverage for a fraction of the credits, and the findings go directly to the developers who can actually fix them.

4. Run a hit list through a workflow

Auditing every page every month is neither affordable nor useful. Most pages don't change month to month, and an unchanged page returns the same findings for the same credits.

A better rhythm:

01 Agree a hit list

Use analytics to pick the pages that matter most: top landing pages, key conversion paths, and content with regulatory or reputational weight. Add one page per template.

02 Let workflows do the work

In Mark, a workflow can step through a slice of that list each month on a schedule, passing each page to the right agents in sequence. Nobody has to remember to press a button, and consumption is more predictable from the length of the list after a couple of runs.

03 Re-run only with a reason

Run a page again when it's changed, or when a fix has gone live. A re-run after a fix is the most persuasive report you can give stakeholders: an issue was there, now it isn't.

04 Keep a human sign-off step

Workflows can pause for approval before results go anywhere, which is where review fits naturally (see point 7).

Working through a list of 20 to 25 priority pages a few at a time covers each one two or three times a year, which is enough for most estates.

5. Trim what each agent does

Once you know what you need from an agent, remove anything that doesn't serve that purpose. Three settings make the biggest difference:

Output steps: Some agents finish their analysis, then spend further credits producing exports in PDF format, sending emails or formatting the same findings twice. If you only need the report, remove the rest.

Reasoning level: Agents can be set to think more or less deeply. A checklist-style review rarely needs the deepest setting; a judgement call on content quality might. Match the level to the task. Test a few settings to see where it makes a difference.

Tools: Every tool an agent can call is something it may decide to use to fulfil its task. Reading web pages and downloading files are typical examples. Limit an agent to the tools its job requires and it's less likely to spend credits on work that doesn't help.

None of these change what the agent is for. They just stop it doing work you didn't ask for.

6. Give agents context instead of making them research

This is the biggest lever of all

A ready-made agent arrives knowing nothing about your organisation, so every run starts with research: what the brand is, what the products are, which rules apply, how the site is put together. You pay for that discovery every time, and the agent can still get it wrong.

Mark lets you hand that knowledge over once. Skills hold reusable instructions and reference material, such as brand rules and tone, product facts, regulatory references and your site's structure, and agents draw on them whenever they're relevant. In practice, these are often simple markdown documents that your team already has, though perhaps not yet in the right format.

The result is an agent that spends its credits on assessing your pages, not on learning about your business. It's also more accurate, because it's checking against your facts rather than its best guess. You also control how that context is shaped and kept up to date.

This is where ready-made agents start to become your agents, and it's the focus of the next post in this series.

7. Build review into the process

In part 1, most of the agents produced at least one incorrect finding that a reviewer would have caught. Efficiency isn't only about credits: a wrong finding that reaches a stakeholder costs far more to unpick than it cost to generate.

A short review checklist

A short, consistent review checklist catches most problems in minutes:

  • Do the headline claims match the detail in the report?
  • Could the agent actually see the page and its assets, or did it work around a block?
  • Are figures, dates and recommendations current?
  • Has any generated code, such as schema, been checked against the live page before it's deployed?
  • Is a lab measurement being presented as if it were what real users experience?

Build this step into the workflow itself, so nothing is published or ticketed until someone has signed it off.

Before your next agent run: a checklist

☐ Each agent answers a specific business question

☐ Every check has one owning agent, and the others are told to skip it

☐ Structural checks run once per template, not per page

☐ A prioritised hit list is running through a scheduled workflow

☐ Pages are only re-run after a change or a fix

☐ Unnecessary exports, tools and reasoning depth are switched off

☐ Brand, product and other relevant context is supplied once, not rediscovered on every run

☐ A human signs off results before they reach stakeholders

What's next

None of these steps needs a big project. Each one is a setting, an instruction or a decision about what to run and when. Together, they turn a set of agents that check everything into a routine that checks what matters, at a cost you can plan for.

Part 3

Why the best results come from a suite of agents tailored to your site, sector and priorities, and the full case for doing this in Mark rather than a general AI tool like ChatGPT, Claude or Gemini.

If you'd like a second pair of eyes on the agents you're running now and where your credits are going, talk to our team. We'll help you find the quick wins.

Not sure where your team stands with agentic AI yet? Take our Agentic Readiness Assessment and see how you compare.

We make technology work for everyone

services_07

Don't forget to share this post!

Jon Seal
Jon Seal
Strategy director

Related posts

AI Optimizely DXP

Opticon 2026 and the Optimizely OMVP Summit: agentic futures, upgrade paths and growth

Oct 16, 2026, 10:00:41 AM
Stephen Gillespie
Optimizely Digital Strategy Digital Roadmap DXP CRO / Experimentation

What is Optimizely. A guide for Digital Leaders

Nov 3, 2025, 2:45:10 PM
Andy Deakin
AI Optimizely

What is agentic marketing?

Sep 8, 2026, 11:38:04 AM
Pam McGee