Key takeaways
- Mirok runs on mirok.ai with the same agents and the same validation protocol as a client account: nothing ships without human approval.
- Three opportunity families dominated month 1: strong pages never cited by LLMs, intent cannibalization on the blog, and technical fixes that are invisible but blocking.
- Every recommendation in the cockpit is audited and sourced before it's proposed, so you can reject an action without losing the diagnosis.
- Fixes that serve classic SEO and LLM citation at the same time deliver the best return for the effort involved.
- The report is published every month to make progress measurable, not to tell a one-off success story.
Why Mirok audits itself every month
Mirok audits itself because a visibility tool that doesn't apply its own recommendations has no credibility. Since month 1, the platform has been running on mirok.ai with the same agents, the same connections and the same validation protocol we deploy on a client account. No special settings, no exceptions: what you read here is what the cockpit produces on a real site, ours.
This report replaces a marketing demo with a verifiable demonstration. Inside you'll find three opportunities surfaced by the SEO, GEO and Copywriter agents, the sourcing behind each recommendation, the fix we chose, and then what was executed and what was rejected. Cockpit screenshots accompany each case to show the actual state of the interface at the moment of the decision.
The value of this format comes from repetition. A demo shows what works once, under hand-picked conditions. A monthly report shows what gets detected, what gets fixed and what stalls, month after month, on a site we know inside out. It's also the only honest way to talk about AI visibility: answer engines move fast, and a single result proves nothing.
The setup: which agents, which sources, which analysis window
The month 1 scope rests on three agents running in parallel on mirok.ai. The SEO agent worked on keywords, page structure and internal linking. The GEO agent tested the site's presence in generative answers and compared the formats competitors get cited for. The Copywriter agent analyzed existing content blocks to flag the ones that answer real questions poorly.
On the source side, the platform is connected to Search Console for queries and positions, Google Analytics 4 for on-page behavior, and the CMS to read and prepare changes. The observation window covers the past month, with a comparison history over earlier periods to tell a trend apart from a one-off blip.
Every recommendation follows the same path before it appears in the cockpit: detection, audit, sourcing, then proposal. Sourcing means tying the opportunity to observable evidence, a test query, a competitor format that gets cited, a crawl data point. You can see how that framework works in detail in the human validation protocol that governs every fix. Nothing goes to production without explicit approval.
Opportunity 1: strong pages missing from LLM answers
Picture a marketing team noticing that a comparison page drives solid organic traffic, yet never shows up when you ask an AI assistant about the same topic. That's exactly the case detected on mirok.ai in month 1: a page ranking well on Google, invisible in generative answers.
The GEO diagnosis isolated three causes. The page answered the question across several long paragraphs, with no standalone answer block up top. It offered no short, citable definition, nothing a model could lift as-is. And the named entities (product name, category, use case) appeared late in the text, muddying the association between the page and the topic.
Sourcing confirmed the problem: across a set of test queries phrased as buying questions, the page was never picked up, while shorter, better-structured competitor pages were. The fix we applied was to move a direct answer block to the opening, add citable definitions, and bring the key entities into the first paragraphs. This kind of restructuring follows a logic documented in the content formats LLMs actually cite, and the result is measured on test queries, not on a general impression.
Opportunity 2: cannibalization and mixed intents on the blog
Take the case of two articles published a few weeks apart that rank for the same query and steal clicks from each other without anyone noticing. On a blog that publishes regularly, this overlap sets in quietly: each page stays visible, but neither consolidates its position.
The SEO agent detected the overlap by cross-referencing Search Console queries with the search intent of each URL. Two articles shared the same primary intent, with wording variations that created the illusion of two distinct topics. Sourcing showed the two pages alternated in the results from day to day, the classic sign of internal competition rather than complementarity.
The call we made was to merge: the more complete piece was kept as the reference page, the other was redirected, and both the H1 and the intro were rewritten to cover the intent explicitly. The rule is simple: one intent, one page. This principle echoes the findings in 500 Mirok audits of B2B SaaS sites, where scattered intent shows up as one of the most common brakes on growth.
Opportunity 3: technical fixes that are invisible but blocking
A site that's fast and well designed for humans can still be misread by crawlers because structured markup is missing on key pages. That's the third case this month: nothing visible to the eye, but signals absent exactly where engines and models expect them.
The Coding agent flagged several points on mirok.ai: incomplete structured data on product pages, internal links that didn't point to strategic pages, and a few markup elements that didn't match the displayed content. Taken one by one, each point looks minor. Combined, they degrade how machines understand the site.
What makes these fixes interesting is their double effect. Clean markup and coherent internal linking improve classic SEO and make it easier for answer engines to pick up the content. That's the heart of the topic covered in the 7 fixes that serve both GEO and technical SEO. Before any execution, each change was presented in the cockpit with its rationale and scope, then approved manually. The agent prepares, the human decides.
What the cockpit executed, and what was rejected
The month 1 scorecard reads in four columns: opportunities detected, approved, executed, rejected. The agents surfaced more opportunities than a team can process in a month, which is precisely the cockpit's job: sort, prioritize, explain.
Some recommendations were approved and executed, mainly the technical fixes and page structure adjustments. Others were set aside for three reasons. Priority first: some actions were relevant but less profitable than others in the same period. Tone second: a few rewrite suggestions didn't match the brand voice. Risk third: changes touching sensitive pages were postponed until their impact could be verified.
That sorting is the proof that autonomy stays under human control. Rejecting an action doesn't erase the diagnosis: the opportunity stays in the cockpit, documented, and can be picked up later. That separation between detection and execution is what makes the system usable on a live site.
The 3 lessons from month 1
First lesson: fixes that serve both classic SEO and LLM citation deliver the best return for the effort involved. Markup, page structure, answer blocks, these are the projects to open first, because they work both angles at once.
Second lesson: intent cannibalization takes time to fix, but it's expensive for as long as it lasts. Merging, redirecting and rewriting calls for editorial judgment, not just technical work. This is the kind of project where human validation adds the most value.
Third lesson: some things are hard to measure in the short term. How citations shift in answer engines depends on refresh cycles we don't control. The prioritization grid that follows is simple: handle what's sourced and reversible first, measure after, and never confuse a lack of immediate proof with a lack of effect.
How to read the next reports
This report opens a monthly series. Each edition will follow the same sections so they can be compared: opportunities detected, fixes applied, rejections with reasoning. Continuity is what gives the format its value, because it turns isolated observations into a trend line.
Month over month, three things will be tracked as priorities: how citations shift in answer engines on test queries, new technical fixes surfaced by the Coding agent, and competitive tracking on the topics where mirok.ai wants to be cited. That last point runs on the same mechanics as the weekly competitor tracking report for LLMs, applied to our own site.
To get the next one, the simplest path is to connect your own environment and compare. Connecting your marketing stack to Mirok in 30 minutes is enough to run a first audit and see the opportunities the agents surface on your site.
FAQ
What does Mirok actually do on a site?
Mirok analyzes your site, detects visibility opportunities on Google and in generative AI answers, then sources each recommendation before proposing it. Prepared fixes only run after human approval. The platform covers technical SEO, GEO, keywords, content, paid and competitive tracking.
Why publish an audit report of your own site?
Because a real-world demonstration, using the same agents clients use, beats a marketing promise. The report shows the opportunities actually detected on mirok.ai, their sourcing and the fixes applied. Readers can judge the evidence rather than a hand-picked screenshot.
Are the fixes applied automatically?
No. Every opportunity is audited, sourced and submitted for approval before execution. Autonomy covers detection, analysis and preparing changes, not pushing them to production. The human keeps the final call, and a rejected recommendation stays documented in the cockpit.
Which agents were used for this first report?
The SEO, GEO and Copywriter agents, connected to Search Console, Google Analytics 4 and the CMS. The analysis window covers the past month, with a comparison history to tell a trend apart from a one-off variation. Other agents, such as Coding or the community agents, will be brought in for future reports.
How often is this report published?
Every month, as a recurring series. Each edition follows the same sections so reports can be compared with one another. Tracking focuses in particular on how citations shift in answer engines, new technical fixes and detected competitive moves.