AI Was Lying About Our Brand. So We Built a Tool to Stop It.

The moment we realized we had a problem

In October 2025, Google Search Console sent us a wake-up call. Pages were not being indexed. Crawl errors were piling up. Our organic traffic had dropped 33%, and we wrote an entire blog post about it.

But indexing issues were just the surface. The real problem revealed itself when we started asking AI assistants about our own company.

We asked ChatGPT, Perplexity, and Gemini: "What does Pixelmojo do?" The answers ranged from wrong to nonexistent. One model confidently described services we had never offered. Another did not mention us at all.

We were not just invisible. AI was making things up about us.

Phase 1: Fixing what Google told us was broken (October 2025)

The git history tells the story. Our first commits were pure triage:

These fixes helped Google. But they did nothing for AI. That required a completely different approach.

Phase 2: Discovering AI could not find us (November 2025)

Once Google's issues were stabilized, we turned to the question that would change everything: can AI systems actually access our content?

The answer was no.

Our robots.txt had no rules for AI crawlers. We had no llms.txt file. Our structured data was a mess of duplicates and fabricated ratings. AI systems had no reliable way to understand what Pixelmojo actually was.

Commit deef381 on November 7, 2025 is when the real work began. We discovered that AI systems were not just ignoring us. They were actively misrepresenting our services. The structured data on our site was sending conflicting signals, and LLMs were filling in the gaps with hallucinated information.

That single commit touched 30 files. It was not a patch. It was a ground-up rebuild of how our site communicated its identity to machines.

Phase 3: Building the playbook by being the guinea pig (February 2026)

By early 2026, we had a working hypothesis: AI visibility is not one thing. It is a stack of interconnected signals that all need to be correct simultaneously. Miss one layer and the whole thing breaks.

We spent February systematically auditing and rebuilding every layer:

Structured data cleanup:

Knowledge graph and entity architecture:

We documented the entire knowledge graph journey in a dedicated blog post. The short version: we built an 18-entity knowledge graph that maps relationships between our services, tools, blog posts, and concepts. LLMs started citing us more accurately almost immediately.

Content optimization at scale:

We did not optimize one post at a time. We overhauled 21 blog posts in a single session, adding StatBlocks, BlogTables, FAQ schemas, speakable markup, and front-loaded definitions to every post. The goal: make every piece of content independently citable by AI systems.

Bot access and crawl policy:

We discovered that blocking AI training bots while allowing retrieval bots actually improved our citation rates. Counter-intuitive, but the data was clear. The commit trail shows the iteration:

Phase 4: The manual process was killing us

By late February, we had a working playbook. But executing it was brutal. Every audit required:

1. Manually checking robots.txt for every AI crawler variant
2. Validating llms.txt format and content
3. Running our domain through 4 different LLMs to check for hallucinations
4. Auditing structured data with multiple tools
5. Checking crawl access with different user agents
6. Comparing what each AI said about us, side by side

Each full audit took 2 to 3 hours. We were doing it weekly. The spreadsheets were getting unwieldy. We started making mistakes, missing regressions, losing track of what we had already fixed.

So we did what any engineer would do: we automated it.

Phase 5: From scripts to tools to Radar (March 2026)

The first tool we built was the AI Crawl Checker. It started as a script we ran locally. Then we realized other people might want it:

Each tool solved one piece of the puzzle. We spent two weeks after the individual tools were built integrating them into a single parallel scan architecture. The real insight came during that integration: the interactions between layers matter more than any individual layer.

A site with perfect robots.txt but broken structured data still gets hallucinated. A site with perfect schema but no llms.txt still gets ignored by some models. You need to check everything at once and see where the conflicts are.

That is why we built Radar.

Phase 6: 50 strangers proved we were not alone

We opened Radar to the public expecting maybe 10 signups. As we covered in Part 1 of this series, 50 people ran full audits in the first week.

The data from those 50 users told us everything:

They were not tire-kickers. As the Part 1 engagement data shows, they averaged over 10 minutes per session. They ran multiple scans. Users came back within 48 hours after implementing fixes to re-scan and verify improvements. When tokens expired, users joined the waitlist instead of leaving. They wanted more access, not less.

The cross-tool conflict detection was the feature users mentioned most. No other platform surfaces the contradiction between what your robots.txt allows and what your server actually blocks. That single insight justified the entire platform for most users.

The problem we solved for ourselves was not unique. Every business with an online presence is about to face the same question: what is AI saying about me, and is it correct?

What 516 commits taught us

Six months. 516 commits. Every mistake documented, every fix timestamped, every iteration preserved in git history.

Here is what we learned:

AI visibility is not SEO. Traditional SEO optimizes for ranking algorithms. AI visibility optimizes for language model comprehension. Different inputs, different signals, different failure modes. You can rank #1 on Google and be completely invisible to ChatGPT.

The stack is interdependent. Fixing your robots.txt without fixing your structured data is like unlocking the front door but leaving the lights off. AI crawlers can get in, but they cannot understand what they find. Every layer depends on every other layer.

Manual audits do not scale. We are a small team and we could barely keep up with our own site. Businesses with hundreds of pages and multiple product lines cannot do this by hand. The tooling gap is real.

Scores must correlate with outcomes. We went from 0 AI citations in October 2025 to being cited by all four major LLMs by March 2026. We know Radar's scoring works because we used it on ourselves, implemented every recommendation, and watched those citations appear. That is not a theoretical claim. It is 516 commits of receipts.

Here is what the before and after actually looks like across all four providers:

The tool we wished existed

Radar exists because we needed it first. Not because we saw a market opportunity. Not because an investor told us AI visibility was hot. Because we were invisible to AI, we were being misrepresented, and the manual process of fixing it was unsustainable.

Every feature in Radar maps to a problem we hit personally:

If you are wondering whether AI is making things up about your brand, it probably is. We know because it happened to us.

Start your own audit

You do not need 6 months and 516 commits. Radar runs the same 12 checks we did manually, in 60 seconds, for free.

• Check your brand free: Run a full AI visibility audit on your domain
• AI Visibility Strategy: Let us implement the fixes for you
• Contact us: Discuss your AI visibility needs

What This Series Covers

This is Part 2 of The AI Visibility Stack, a five-part series documenting how we built an AI visibility platform from scratch. Each post builds on the last, from user validation to founder story to product deep-dives.

All 5 parts published.