I haven’t written a Building Toket note for a while.

Not because there was nothing to write about. If anything, I’ve been researching too many things at once.

I started with GEO: trying to understand why AI mentions one brand but not another, why a brand gets mentioned but not recommended, and whether AI actually understands what a website is saying.

Then I somehow ended up researching AI infrastructure.

GPUs, electricity prices, data centers, liquid cooling, rack power density, deployment conditions across different countries, CAPEX, OPEX, and payback periods.

On the surface, this has very little to do with GEO.

But after working through all of it, I realized I had been trying to solve the same problem all along:

How do I know whether the answer in front of me is actually true?

AI Can Give You an Answer Very Quickly

That is the easy part now.

Ask a model:

Which country is better for deploying AI infrastructure?

It can quickly tell you about power stability in Japan, electricity-cost advantages in parts of Northern Europe, mature infrastructure in South Korea, or abundant hydropower in Paraguay.

Keep asking:

How much does a B300 cost? How much power does a GB300 NVL72 consume? What is the industrial electricity price in France? How long would it take an AI data center to pay back its investment?

It will answer those questions too.

It can even turn everything into a nice comparison table.

The problem begins when I actually try to put those numbers into a product.

Prices for the same hardware can vary significantly across sources.

Electricity prices can mean residential or industrial rates, with or without taxes, at different consumption levels, and across different reporting periods.

Some numbers come from governments. Some come from vendors. Some are quoted by the media. Others have been copied across so many websites that it becomes difficult to find the original source.

At that point, “AI knows the answer” becomes much less interesting.

A more important question is:

Where did this number come from?

A Number Should Come With Evidence

Eventually, I gave myself a simple rule.

If a number has no source, I don’t treat it as a fact.

If I don’t know, I show that I don’t know.

If something is only an estimate, the product should clearly say that it is an estimate.

Hardware pricing is a good example.

You can find public price ranges for B300 systems. That does not mean NVIDIA has published an official standard price for a single B300 GPU.

The GB300 NVL72 has a similar problem.

Public quotes and industry estimates exist, but when the sources themselves disagree significantly, the worst thing a product can do is pick one number and pretend that it is precise.

So I started adding more and more things that once felt unnecessarily complicated:

Sources, dates, regions, units, reliability, and verification status.

Sometimes even an individual field has its own sources.

A few years ago, I would probably have considered all of this overhead.

Now I think it may be where much of the actual value is.

The calculation itself is not that difficult.

The difficult part is deciding:

What deserves to be included in the calculation?

GEO Has the Same Problem

This brought me back to something I had already encountered while building GEO.

Imagine a brand’s AI visibility moves from 9.7% to 18.8%.

That looks like a very clear improvement.

But keep asking questions and things become complicated very quickly.

Did the brand actually improve?

Did negative content disappear?

Did search results change?

Did the model change?

Or were the prompts, retrieval sources, and platform behavior simply different from the previous test?

If you only look at the final percentage, it is very easy to tell a convincing story.

I’m becoming less comfortable doing that.

AI platforms are black boxes.

The result we observe today does not automatically explain why that result happened.

That is why I have gradually started paying much more attention to samples, sources, raw responses, and retest conditions in GEO.

Not because those things make a report look more professional.

Without them, the percentage itself is not as reliable as I once thought.

AI Is Making Generation Cheap. Evidence Is Still Expensive.

This is probably the biggest thing I’ve noticed recently.

Research products used to spend a lot of time organizing information and producing the final output.

AI can now do much of that in minutes.

Writing is fast.

Summarization is fast.

Translation is fast.

Classification is fast.

Even writing code can be fast.

But verifying a fact is still slow.

You still have to find the original source, understand the methodology, check the date, confirm the units, and make sure the number has not been misquoted somewhere along the way.

Sometimes I spend longer verifying one number in a dataset than AI spends generating an entire page of the product.

There is something interesting about that.

AI has made answers cheap.

It has not made facts cheap.

If anything, the opposite may be happening.

When the cost of generating an answer approaches zero, deciding whether that answer deserves to be trusted becomes more important.

It Is Changing How I Build Products

I used to start with a simple question:

Can this feature be built?

Now I increasingly start with another one:

What gives the product the right to reach this conclusion?

If I cannot answer that question, I would rather show “No reliable data available” than manufacture certainty.

That may make the product look less intelligent.

But it may also make it more honest than an AI product that always has an answer.

I don’t know whether this is ultimately where Toket is heading.

I’m still working on GEO. AI infrastructure is just the latest rabbit hole, and I’ll probably end up researching many other things that look completely unrelated.

But one thing is becoming increasingly clear:

I’m not very interested in building another AI product that simply gives people answers.

I’m becoming much more interested in the evidence behind those answers.

And when the evidence isn’t good enough,

maybe the product should be able to simply say:

I don’t know.

Estimate task cost in the AI Cost Analysis or refine prompts in the Prompt Optimizer.