Why We Built Radar: The Commit Trail Behind Our Own AI Visibility Fixes

Why Did Pixelmojo Build Radar?

We built Radar because checking our own site for AI visibility problems by hand did not scale. The git history from October 2025 to March 2026 records what we changed and when. It does not record what AI engines said before and after, so this version of the story sticks to what the record shows.

The first version of this post, published on 5 April 2026, told a cleaner story: zero citations, a 30-file fix, then citations from all four major engines. When we checked it against the repository on 2 October 2026, several of those details were wrong, and the central claim had no record behind it. This rewrite keeps the parts we can show and says plainly which parts we cannot.

What Was the First Sign Something Was Wrong?

The first sign was Google Search Console, not an AI engine. In early October 2025 Search Console reported pages that were not being indexed and URLs that returned 404 errors, and our first commits of the month were triage.

These were search engine fixes. They made the site easier for Googlebot to crawl and index. They said nothing about how an AI assistant would describe the business, which is the question that came next.

On 31 October 2025 we added explicit rules for AI crawlers to robots.txt (commit 8ae6b225) and, the same day, rules for their user-agent variants (982e8e77). Before that, the file had no AI-specific rules at all. The same day we restructured our llms.txt so it was easier to parse (6b21b5d7).

What Did an AI Engine Get Wrong About Us?

On 7 November 2025, Perplexity described Pixelmojo as not offering two services we did offer. We know this because the commit that responded to it says so in its message.

“Problem: Perplexity incorrectly shows PixelMojo as "No" for full-stack development and digital transformation.”

Message of commit deef3810, 7 November 2025

The commit changed six files and added 462 lines:

• it expanded public/llms.txt with full-stack and digital transformation sections (the file already existed, added on 28 September 2025 in commit 9131e818);
• it added a 273-line AI capabilities factsheet in public/;
• it listed more than 20 capabilities in the site's structured data;
• it rewrote meta descriptions to mention backend, databases and infrastructure;
• it let crawlers fetch the new factsheet in robots.txt.

The commit message closed with an expected impact: AI systems would now represent the full range of services correctly. That line was a hope. Nobody recorded what Perplexity said a week later, or a month later. So the commit is good evidence of the problem we saw and of what we tried. It is not evidence that the fix worked.

What do we know now that we did not know then?

In November 2025 we treated llms.txt and structured data as levers that would change what AI engines said. The evidence since then points the other way.

SE Ranking's study found that removing llms.txt from its prediction model made the model more accurate. Google's guide to its generative AI features, published in May 2026, lists llms.txt under the things site owners do not need to do. We cover both in Google Says You Don't Need llms.txt.

“There's no correlation between AI citations and LLMs.txt.”

SE Ranking, LLMs.txt research, 7 November 2025

That does not make the November commit pointless. Accurate structured data and a clear description of services are good hygiene for people and search engines. What it does mean is that we should not have described those files as the reason anything changed in AI answers.

What Did We Change Between November and March?

From February 2026 we cleaned up structured data we should never have shipped, built a knowledge graph of our own entities and fixed crawl problems Search Console kept reporting. The commit dates below come straight from the repository; several dates in the earlier version of this post were off by days or weeks.

The first row deserves a plain statement. Our site carried aggregate rating markup that no real reviews supported, and we removed it on 15 February 2026. Fabricated ratings are a problem whatever they do to AI answers, and removing them is the kind of fix we would make again today.

The knowledge graph started on 24 February 2026 and lives in code, in src/lib/knowledge-graph.ts. As of 2 October 2026 its validator checks 61 entities. We wrote about how it was built in Knowledge Graphs and LLM Visibility.

The 27 March commit added speakable schema to our articles, linked our organization to the knowledge graph and retrofitted 21 blog posts with 24 StatBlocks. Part of that work is a dated note of its own, because the original post presented formatting as a reason AI engines cite a page.

Why Did Manual Checks Turn Into Tools?

Because we kept doing the same checks by hand. For each round we read robots.txt crawler by crawler, checked llms.txt, reviewed structured data, fetched the site as different bots and asked several assistants what Pixelmojo does. Each check was simple. Doing all of them, repeatedly, and comparing the results was not.

The first tools were the checks we ran most often:

Two of those commits are about the crawl checker getting things wrong. Web application firewalls blocked our test requests, and an early version judged bot access by HTTP status codes instead of by the site's stated policy. Both are the kind of mistake you only find by running a tool against real sites, and both shaped how Radar reports a check it cannot complete: it marks the result withheld instead of grading it.

How Did the Tools Become Radar?

The tools became one audit because their results mean more together than apart. A crawl check that passes while the citation check finds no mentions points to content, not access. A robots.txt that allows a bot while a fetch as that bot fails points to the server, not the policy. Radar's cross-tool insights read those combinations from the results of a single audit.

The milestones after launch matter as much as the ones before it:

• 17 March 2026: the Radar platform commit (452c6da4) put the tools behind one audit.
• 31 March 2026: all 12 tools of the time ran by default (d280479f).
• 19 April 2026: Radar launched publicly with a free first check and paid audits.
• 5 June 2026: Brand Disambiguation became the 13th tool (585df483). Posts in this series written in April that said "13 tools" were ahead of the product.
• July 2026: a customer reported an audit that had measured a different company with the same name. Part 6 covers that failure and the entity resolution we rebuilt after it.
• 12 August 2026: audits began running their tools in staged batches, so provider-heavy checks run on their own.

What Can the Commit Trail Prove, and What Can It Not?

A commit trail proves what changed in the code and when. It cannot prove what changed in an AI engine's answer, because the answer lives outside the repository. The table below sorts each claim from the first version of this post by what the record supports.

The last point in that diagram is why we do not try to reconstruct the missing evidence now. AI answers change between runs, even for the same question. A September 2026 study of product queries found that the products recommended often changed across repeated requests. A single answer recorded today would tell us little about October 2025, and nothing about which of our changes mattered.

What Is Radar Today?

As of October 2026, Radar runs in two steps. A free check runs 6 technical readiness tools on one domain, once per domain, and does not ask AI engines about your brand. A paid full audit runs all 13 tools, including Reddit monitoring and the tools that sample answers from ChatGPT, Claude, Gemini and Perplexity, and costs $5 for one audit, $12 for three or $40 for ten. Pro is $199 a month for 40 audits with watched domains.

Read the transcript

I'm Lloyd, founder of Pixelmojo. We build AI products for discovery, lead qualification, and business operations. Radar is where many teams start. It helps you inspect how AI describes your business, examine the evidence, and decide what to improve next. Your team makes the changes, then checks what happened. We're also developing Radar Sites, a visual builder for websites. Vector qualifies and routes incoming leads around your sales process, giving your team a clearer next step. Hive coordinates specialized AI co-workers so they can share context, hand off work, and bring in people. With Vector and Hive, you own the custom application. We maintain the intelligence. Welcome to Pixelmojo.

Radar does not change your site. It shows what it checked and what it found, ranks what to look at first and helps you draft a fix, and your team makes the change. The Radar user guide walks through the free check, the paid audit and the dashboard step by step.

“Your team makes the changes, then checks what happened.”

Lloyd Pilapil, Meet Pixelmojo (video, September 2026)

What Changed in This Post?

We rewrote this post on 2 October 2026 as part of a review of every post on the blog. The changes:

• commit dates, hashes and file counts now come from the repository;
• the zero-to-cited claim and the before-and-after AI answers are removed, because nothing recorded them;
• the "516 commits" line is now 515 commits of every kind;
• the November fix is described as what we believed at the time, alongside what the evidence says now;
• the claim that 50 people ran full audits in the first week is replaced by what our records show (Part 1);
• Radar is described as it runs today, with 13 tools since June 2026.

What This Series Covers

This is Part 2 of The AI Visibility Stack, a six-part series on how we built Radar and what we learned running it. Part 1 is a dated note on the first beta weeks. Part 3 covers why SEO tools and AI answers need different measurement.