We ran seven Mark agents against real sites. Here's what's worth your credits.
Getting the most from Mark agents · Part 1 of 3
Optimizely's Mark agents can audit your site every month without anyone pressing a button. We tested seven of them on live sites to find out which ones earn their credits, which ones overlap, which need significant tailoring before they're truly useful, and which still need a human looking over their shoulder.
7
agents tested on live sites
22–200
credits per agent run
200
free credits a month per instance
6 of 7
agents produced at least one erroneous finding
Most organisations review their websites in bursts. A big audit usually happens when something goes wrong: a complaint, a failed accessibility check, a sudden drop in traffic. Fixes get made, attention moves on, and the site drifts until the next problem forces another look. It's rarely a culture of steady, proactive improvement.
Agentic AI changes that. With Optimizely's Agent Platform (formerly Opal, and now home to its AI, "Mark"), you can line up a list of pages and have a workflow push them through specialist agents on a schedule. Accessibility, AI search readiness, content quality and performance can be checked every month, not just when something breaks.
Agents can also check pages before they go live, but in this series we're focusing on the existing estate: steady, incremental improvement to the pages you already have.
That's the promise. We wanted to see how much of it holds up, and what it costs in time and consumption credits.
How we tested
We took seven of the ready-made agents from Optimizely's agent directory and ran each one several times on live pages across financial services, utilities, manufacturing, membership bodies and charities:
GEO Auditor: how ready a page is to be found and quoted by AI search tools
EEAT Checker: the experience, expertise, authority and trust signals Google looks for
Web Accessibility Evaluation: a WCAG 2.2 review plus Lighthouse scores
Image Alt Text Generation: reviews and rewrites image descriptions
FAQ Creation: researches common questions and writes an FAQ section with schema
Page Performance Evaluation: Lighthouse and real-user speed data with a fix list
Competitor PageSpeed Analysis: your page's speed against a rival's
For each run we recorded time taken and credits used, then read every report as a client would: is this accurate, is it new, and would anyone act on it?
It's a small sample, so the numbers are more of an indicative guide than a benchmark.
Worth knowing
One site we reviewed blocked the agents entirely, and that's becoming more common. Unwanted bot traffic drives up hosting costs and skews analytics, so some businesses now block bots wholesale. That solves real problems, but a blanket ban also locks out the agents you want working on your site. If you plan to use automated audits, let the agents you trust in while keeping the rest out. There's a whole other topic we'll explore around defining the right AI strategy for your business.
What impressed us during the trial
The best agents found things a busy team would miss, and found them in minutes.
A page title that contradicted the page
On a manufacturer's product range page, the GEO Auditor spotted that the title tag quoted a weight range that didn't match a single product listed below it. To a person skimming, it's invisible. To an AI search tool deciding whether to trust and quote the page, it's a red flag. The same run found product specs that disagreed between the product cards and the FAQs further down.
Two different numbers for the same claim
On a campaign page, the EEAT Checker noticed the meta description and the body copy quoted different percentages for the same headline benefit. It also flagged named competitor comparisons with no published test method, which is a brand and advertising-standards risk as much as an SEO one. That run cost 50 credits and took two minutes.
No structured data at all
On a utility's main service page, the GEO Auditor found no structured data (schema) of any kind, then wrote ready-to-review code to fix it.
Leftovers from development
The accessibility agent found placeholder text from a test build still showing in a live breadcrumb trail, along with a "skip to content" link that didn't go anywhere.
None of these are exotic. They're the everyday slippage that builds up on any large site. The value is in catching them routinely, before a customer, a regulator or an AI search engine does.
Where the credits go
Agent runs cost between 22 and 200 credits each, and value bore little relation to cost.
Consumption isn't a fixed property of the platform. It depends on how tightly each agent's instructions are written, which tools it can call, how much context it has to digest, and the level of reasoning it's set to use. The ready-made agents are built to work for anyone, so they tend to err on the side of doing more.
At time of writing, Optimizely gives every instance 200 free credits a month. One decent accessibility review used all 200, and a GEO audit used 90. The free allowance covers about two page audits a month: enough to try the agents, but certainly not enough to run an optimisation programme.
A bigger issue is overlap
Each agent is built to stand alone, so each one checks the basics for itself:
- The GEO Auditor runs its own performance tests, so a separate performance agent largely repeats it.
- The accessibility agent reviews image descriptions, and so does the alt text agent.
- Structured data, headings and meta descriptions came up in three different agents' reports.
Run the full set on the same page and you pay several times over for the same findings. That's fine for a test. As a monthly habit across dozens of pages, it adds up quickly.
Some of the cost is avoidable within a single agent too. The accessibility agent produced its main report early in the run, then spent the rest of its time on multiple PDF export attempts and sending an email. Strip those out and the same insight should cost noticeably less. We'll cover this and other ways to optimise consumption in part 2.
Where it still needs a human
Every report reads confidently, which is exactly why each one needs checking. Six of the seven agents produced at least one erroneous finding that a consultant or client reviewing the report would have caught.
Filling gaps when it can't see
One agent noted that the site had blocked it from viewing the images, yet still produced a full set of filenames and suggested descriptions based on guesswork rather than what it had seen.
Contradicting itself
One report's summary claimed a lead on four of six measures, while its own table showed a lead on two.
Out-of-date advice
Another agent recommended watching for a search feature that Google restricted in 2023, and repeated figures from the page without checking they were still current.
Lab numbers that scare people
Some reports quoted lab load times of up to 30 seconds next to real-user data showing about one second. Without context, the lab figure is the one a stakeholder remembers.
None of this makes the agents unusable. It means a person who knows the site and the subject should review each report before it reaches a decision-maker.
The agents do the legwork; people still decide what's true and what matters.
Our verdict
Two agents definitely would earn a place in an automated process straight away; the rest would likely need reshaping for each client, and in some cases adopting a slower rhythm or replacing entirely depending on the other choices made.
| Agent | Average credits per run | Verdict | Why |
|---|---|---|---|
| GEO Auditor | 90 | Keep, run monthly | The most specific, useful findings per credit, plus code you can deploy after review |
| EEAT Checker | 40 | Keep, run monthly | Cheapest substantive run; finds brand, trust and compliance risks senior people care about |
| Web Accessibility Evaluation | 200 | Reshape | Finds real defects, but much of each report is generic manual-check guidance. Remove the extras and run it once per page template, not per page |
| Competitor PageSpeed Analysis | 67 | Run quarterly | A useful benchmark for stakeholders; speed doesn't shift enough to justify monthly runs |
| FAQ Creation | 115 | Use as needed | Good content output, but it's a writing tool, not an audit. Use it for priority pages |
| Page Performance Evaluation | 60 | Replace | Largely repeats PageSpeed Insights data already covered by the GEO Auditor |
| Image Alt Text Generation | 42 | Drop as standalone | Largely covered by accessibility checks, and unreliable when it can't see the images |
Why not just use ChatGPT, Claude or Gemini?
This is a reasonable question. If you already have one of those tools to hand then what's the point of investing time and money in yet another AI solution? And for a one-off review of a single page, a general AI tool will do a reasonable job. Mark itself runs on the same underlying models.
The difference is what sits around the model. In Mark, the audit is a workflow that runs on a schedule, works through an agreed list of pages, and draws on brand rules and site knowledge set up once rather than pasted into every prompt. It's logged, shared across the team, and sits inside the Optimizely estate and contract you already have. A chat window waits for someone to type.
We'll make the full case, including when a general tool is the better choice, in part 3.
What's next
Automated site audits with Mark are worth doing, but not by switching on a series of agents and letting them run. The value comes from choosing the right few, adapting them for your context, running them on the right pages at the right cadence, and having someone review the results.
Part 2
The practical ways to get more from every credit: cutting overlaps, auditing templates instead of pages, and giving agents the context they'd otherwise spend credits rediscovering.
Part 3
Why the best results come from a suite of agents tailored to your site and sector, and why that's easier to build in Mark than anywhere else.
If you're already using Mark agents, or wondering whether they're worth the credits, talk to our team. We're happy to look at what you're running and where the value is.
We make technology work for everyone
