A little after seven this morning I had Claude open in a second browser window, watching it work through Google Search Console one report at a time while I enjoyed the sunrise on our desk, what I call our deck shala. Why was I doing this? Because I spend a lot of time writing these BIP posts, between 30 minutes to 2 hours per day. It is a huge time commitment. I also know that SEO takes about 3 months to start working, so before this time, I was not expecting miracles.
I went to click in around Google Search Console, and then thought. Let's have Claude build a Chrome prompt to run through GSC and give a report. Much better than me doing it.
The night before I had written it a prompt: go through every report, read the real numbers off the screen, and hand me back a ranked list of what is broken. It moved through the clicks, the impressions, the pages that rank, the pages that get nothing, and wrote the whole thing up while I sat there doing almost nothing. Call it an ai audit you run on yourself. Twenty minutes of agent time for what used to be an afternoon with a spreadsheet.
Then I pointed the same approach at my analytics.
First, the thing I have been putting off for weeks is digging into this. The scoreboard I grew up on, ranking in Google's blue links, is now one of three. There is SEO, still real, still about ranking in search results. There is AEO, answer engine optimisation, being the source an assistant quotes when someone asks ChatGPT or Claude or Perplexity a question. And there is GEO, generative engine optimisation, showing up inside the answer a model writes instead of in the list of links underneath it. In 2026 the three have started to pull apart. A site can win one and quietly lose the other two. So I wanted my own checked against all three, with ai doing the legwork. An ai audit for startups, run on the person who built the thing.
On the search side, the agent found dumb but expensive stuff. 35 of my 81 page titles were too long and getting cut off in Google before they made their point. 73 of 81 meta descriptions ran past the limit, so the snippet a searcher actually reads was clipped halfway through. The homepage was loading 2.8MB of images, which I brought down to 177KB by converting three hero files to a lighter format. None of it is clever work. It is all the agent prompting me through with suggestions. Easy, anyone can do it.
All of it was costing clicks I had already earned.
The analytics pass is the one that stung. Every marketing page carried my tracking tag. Not one blog post did. For three months I had been publishing daily, and the whole blog, the thing I put the most hours into, was invisible to my own analytics. I had been writing into a room with the counter switched off, then squinting at flat numbers and wondering why. The audit caught in twenty minutes a hole I had not seen in ninety days.
Another embarrassment, and I am putting this out in public as per the discipline.
Then the part most people skip, the AEO and GEO layer. The good news came first. The site already had what an answer engine needs to read it cleanly: an llms.txt file, the plain text page that tells a model what the business is and who it serves, structured data on every page, and clean URLs. Better, there is already one question where the site is the answer and not just a link. These are the tools I have been tweaking daily with the business-to-agent SaaS idea for giveready.org.
Ask about hiring an ai financial controller versus a human one and my page comes back first. That is the AEO win that is already ranking. Being the specific, honest answer to a real question is what gets you quoted.
If you run a site and have never looked at it through an ai agent's eyes, you are almost certainly polishing one of those three and ignoring the other two. Most people are still tuning for 2013 Google and have no idea whether an answer engine can even read them. Maybe that is not true, but it was for me, and I am building systems to change that.
The first lesson is old, and I relearn it about once a quarter. The flat blog numbers were never a content problem. They were a missing counter. You do not get to argue with a number you never collected.
The second is that AEO and GEO are not won with tricks. They are won by being the best, most specific answer to a real question, and then making sure the plumbing lets a machine read you: the llms.txt, the structured data, the clean URLs. Write the thing a person would thank you for, then make it legible to the models, which is most of what being an ai-first company comes down to now. This is classic long-tail search, high intent terms that are built better through good organic writing. I have been open about how I use ai to co-author my writing.
The third is that the prompt is the asset, not the audit. I wrote two careful prompts, one for Search Console and one for analytics, and the agent will run them on any site I point it at, today or in a year. It is the same idea I keep circling back to in this build, thin harness, fat skills: the judgment lives in the instructions I wrote down, and the agent is just the loop that runs them.
If you want to run this on your own site, it goes like this.
One. Point an ai agent at your own Search Console with a written prompt. Make it go report by report, read the real numbers off the screen, and hand back a ranked list of problems, not a tidy summary.
Two. Do the same with your analytics. The very first thing it should check is whether every page actually carries your tag. Mine did not, for three months.
Three. Audit the answer-engine layer. Check for an llms.txt served as plain text, structured data on every page, clean URLs, and a plain-language statement of what the site is and who it is for, so a model can quote you correctly.
Four. Fix the cheap mechanical leaks first, the titles, the descriptions, the image weight. Then pick the one question you most want to be the answer to, and go be the best answer to it.
The whole thing, both search engines and the answer-engine pass, took an afternoon. A few years ago it was a week of someone's time and a five figure invoice. Tomorrow I point the same two prompts at a client's site and watch them work again. Same instructions, new site, no extra effort, which is the closest thing to free money this job offers.
Monthly Revenues $11,000 | Clients 2 | Prospects (AI outbound agent now live) | Team: Me + part time Jan (CTO)
Day 79 of 365.