Niels

Passionate full-stack builder. Dad, dog dad, based in e-Estonia 🇪🇪

[Niels @nielskaspers]

Most teams doing GEO are making the same mistake:

They're trying to “optimize for AI” on their own site.

That solves extraction.
It does not solve recommendation.

My simple playbook:

1. Build pages AI can quote
Comparisons, alternatives, use cases, pricing, original data.

2. Build proof AI can trust
Named experts, real stats, customer examples, public wins.

3. Build mentions you don't own
Reddit, YouTube, review sites, newsletters, partner ecosystems.

Your site tells AI what you do.
The internet decides whether to believe you.

2026 search is less “rank for keyword”
and more “be the obvious answer when the prompt gets specific.”

[Niels @nielskaspers]

The most important AI product shift this week is not another model benchmark.

It's stack compression.

Voice used to mean STT + LLM + TTS + telephony + orchestration.
Now xAI is bundling the whole thing at roughly $0.06/min all-in.

Same pattern with MCP: the interface goes horizontal, the value moves to whoever owns workflow, context, and distribution.

If your startup is one neat layer in the middle, that is not a moat.
It's a line item the model provider is coming for.

[Niels @nielskaspers]

Hot take:

Most teams doing “AI SEO” are optimizing the wrong surface.

LLMs don’t care about your FAQ schema.
They care whether the web already agrees you’re a credible answer.

SEO let your page outrank your reputation.

AEO/GEO flips it:
your reputation outranks your page.

Original data.
Third-party mentions.
Clear category positioning.

That’s the moat.

[Niels @nielskaspers]

Builder lesson from this week's AI product launches:

The winners are getting narrower, not broader.

xAI turned voice into one productized workflow.
Ivo turned negotiation history into contract memory.
Different products, same lesson:

Pick one job users already understand.
Then go absurdly deep on context.

Generality demos well.
Specificity retains.

[Niels @nielskaspers]

GSC adding AI visibility this week killed a lot of fake GEO theory for me.

llms.txt won't save you.
FAQ spam won't save you.
Schema cosplay won't save you.

If AI can't verify what you do and why you're credible, you won't get cited.

It doesn't reward tricks.
It rewards evidence.

[Niels @nielskaspers]

GEO is not a schema problem.

It's a mention problem.

LLMs are citing Reddit, reviews, comparison pages, and docs.

Playbook:
Own 5 prompts.
Publish pages that answer them.
Seed mentions on cited surfaces.
Add proof: stats, pricing, screenshots.

Mentions beat markup.

[Niels @nielskaspers]

Google Search Console finally added AI performance reports.

The timing is perfect:
teams just got a new dashboard for a channel where the click is disappearing.

This week the signals were brutal:
position 1 CTR on AI-heavy queries is getting crushed
AI sessions often end with no click
schema theater still moves almost nothing

Most brands will respond by measuring AI search harder.

Wrong move.

The real job is becoming the source the model trusts enough to quote:
original stats
specific buyer questions
earned third-party proof
clean pages built for extraction

The dashboard arrived right after the click stopped mattering.

[Niels @nielskaspers]

Most founders still treat AI visibility like SEO.

It's closer to product + retention.

Anyone can win one citation with formatting.
Very few can stay recommended.

The models keep re-evaluating:
- Is the answer clear?
- Is there fresh proof?
- Do credible people mention you?

In AI search, distribution is no longer a page.
It's a living evidence loop.

[Niels @nielskaspers]

Builder lesson from this week's agent-memory wave:

If your AI product "needs memory," your first problem is not vector search.

It's deciding:
- what deserves to be remembered
- what must expire
- who is allowed to overwrite it
- what happens when memory is wrong

Most teams are treating memory like infra.
The winning teams will treat it like product behavior.

[Niels @nielskaspers]

New growth playbook for AI products:

1. SEO gets you found.
2. AEO gets you quoted.
3. MCP/API gets you used.

Most teams still optimize for clicks.
The new funnel is:
search result -> AI answer -> agent action

If your product isn't easy to index, cite, and call, you lose before the user ever hits your site.

[Niels @nielskaspers]

The hottest job in AI isn't prompt engineer.

It's the forward-deployed builder who sits inside the workflow, cleans up the ugly edge cases, and gets agents to produce revenue.

AI Engineer World's Fair has a full FDE track now.

Models are abundant.
Deployment is the moat.

[Niels @nielskaspers]

Sharp opinion:

AI product growth is becoming onboarding engineering.

The flashy demo is not the bottleneck anymore.
The drop-off happens right after, when the user has to connect data, grant permissions, import context, and trust the thing with real work.

That means the new growth levers are not just copy or reach.
They're:
fast setup
clear scope
easy undo
visible audit trail

In AI, activation is not "first prompt sent."
It's "first workflow delegated."

The teams that win won't have the smartest model.
They'll have the shortest path from curiosity to trust.

[Niels @nielskaspers]

Builder lesson:

The teams getting real value from AI right now are using it on themselves first.

Not for demos.
For support queues, QA, reporting, onboarding, and internal search.

Why?
Because your own workflow exposes the truth faster than customer interviews do:
- where context breaks
- where permissions block usage
- where the agent sounds smart but can’t finish the job

In AI, dogfooding became product research.

[Niels @nielskaspers]

AI search changed how I work more than how I write.

I spend less time asking:
what should we post?

More time asking:
can a model name the product?
can it find the page to cite?
does that page route a human to the next step?

Content used to be the asset.
Legibility is.

[Niels @nielskaspers]

Everyone is shipping "agents" this week.

My rule: don’t build one unless the workflow has:
1) repeated intent
2) clear action surface
3) verifiable outcome
4) cheap recovery
5) economic upside after failures

Miss one, and you’re shipping a demo dressed up as a product.

[Niels @nielskaspers]

The AI hype cycle is entering its boring phase.
Good.

Google now wants you measuring AI Search impressions.
Agent startups are launching dashboards for usage and cost.
AI Engineer World's Fair opens with evals, infra, and reliability everywhere.

That is the tell.

AI is moving from "look what it can do"
to "prove what it did."

The winners won't be the noisiest model wrappers.
They'll be the teams that can measure behavior, tighten feedback loops, and show ROI without hand-waving.

[Niels @nielskaspers]

Hot take:

the next great growth team will look less like SEO + paid social
and more like merch ops for machines.

Google is turning Search into a shopping agent.
ChatGPT keeps improving shopping.
Google just added generative AI visibility reporting in Search Console.

That means your product page is no longer just persuading a human.
It is feeding answer engines, recommendation agents, and bidding systems.

Beautiful brand copy will matter less than:
structured specs
clear comparisons
real reviews
clean feeds
obvious differentiation

The brands that win AI search won’t be the loudest.
They’ll be the easiest for machines to understand
and the safest to recommend.

[Niels @nielskaspers]

Builder lesson from this week’s AI noise:\n\neverybody wants better agents.\nvery few teams are fixing the context those agents run on.\n\nLevie keeps repeating evals.\nBox keeps repeating context.\nSame lesson.\n\nStale docs, messy permissions, unclear owners, and missing recovery paths are not IT issues anymore.\nThey are product issues.\n\nAI does not just automate the workflow.\nIt stress-tests whether your company is actually operable.

[Niels @nielskaspers]

This week reminded me that AI product quality is quietly moving from answer quality to follow-through quality.

The moment scheduled tasks become native, the question changes:
not "is the model smart?"
but "will it remember, run, and come back with something useful?"

In my own workflows, the AI I keep is rarely the flashiest.
It's the one that shows up on time with context.

Reliability is becoming the product.

[Niels @nielskaspers]

Google finishing its June spam update is your reminder:\nAI visibility is not a prompt hack.\n\nMy playbook:\n1. Publish claims you can prove\n2. Make the proof crawlable\n3. Earn mentions you don’t control\n\nIf a model can’t verify you, it won’t trust you.