On a Saturday morning I sent my AI coworker a long, rambling message about an idea I could not decide if I loved or hated.
The idea: build a landing page for every micro-niche an ecommerce brand could be in. "Email agency for ecommerce handbag brands." "Email agency for pet supplement brands." Dozens of them. Maybe hundreds.
The logic was not crazy. If someone tells ChatGPT they own a handbag brand and then asks for an email agency, the agency whose site literally says "email marketing for handbag brands" should win that recommendation over a generic ecommerce agency.
The other half of my brain said this is 1990s SEO and it will look like spam.
Here is the actual message I sent (this is for an agency I co-own, not a client):
It told me no, and explained the mechanism
What I wanted was a yes. What I got was a yes to the instinct and a no to the plan.
The part that landed for me was the second-order problem. Everyone worries about whether AI will "like" the pages. The real issue is that AI answer engines mostly run a live web search and read what comes back. If Google buries 200 templated doorway pages, the AI never sees them at all. You are not gaming the model, you are just failing the search step.
The three tiers it gave me instead
- Tier 1: 5 to 8 real industry vertical pages. Not "handbags." Fashion and apparel. Health and supplements. Beauty. Food and beverage. Each with genuinely different content: the fashion page talks collection drops and lookbook-driven design, the supplements page talks subscription flows and claims compliance. Then mention the sub-niches naturally inside the page, so the word "handbag" lives on a legitimate page instead of getting its own thin one.
- Tier 2: niche blog posts. "5 Email Flows Every Handbag Brand Needs on Klaviyo." Specific enough for an AI to match, educational enough that it is not spam. Each one internal links up to its industry page.
- Tier 3: third-party mentions. The one I did not want to hear.
I did not want to hear it because tier 3 is the only tier you cannot just build. Pages I can ship on a Saturday. Getting onto somebody else's "Top 10 Klaviyo Agencies" list takes outreach, reviews, case studies, and time. It is the least fun and it moves the needle most, and every agency owner I know is doing tiers 1 and 2 and skipping it.
Six pages live in 35 minutes
I said go. Build the industry pages, add an Industries dropdown to the nav, and pull the real client email examples from our examples page into each one so the pages have actual work on them, not stock art.
It mapped 14 client examples across 6 verticals, told me which two verticals were thin BEFORE building them (one had a single example), then shipped.
Flagging thin pages unprompted is the behavior I actually pay for. It would have been easy to build all six at equal depth and let me discover the weak ones later.
Then I opened the menu on my phone
Broken. The new Industries dropdown would not open, and the font was slightly off from every other nav item.
First fix attempt did not work. I sent another screenshot. It stopped guessing, pulled the live file down, and found the actual cause.
It had written the dropdown JavaScript inside a <script type="application/ld+json"> block. That is a structured data container. The browser reads it as text and never runs it. So the code was perfect and did nothing.
Two things I take from that. One, it broke it. Two, when I pushed back a second time it switched from patching to pulling the real file and diagnosing, which is what a good developer does and what most AI tools refuse to do because guessing is faster.
Last note was mine: it had put a little yellow "By Industry" label above the links. Redundant, since the dropdown already says Industries. That label only earns its spot in the Services menu where there are multiple categories. Gone in one message.
What I'd tell another agency owner
- Specificity does win with AI recommendations. Your instinct there is fine.
- Get there with 6 real vertical pages plus niche blog posts, not 200 find-and-replace pages. The doorway version fails at the search step before a model ever reads it.
- Put your sub-niche words inside legitimate pages. One good page can hold ten sub-niches.
- Schema listing the industries you serve does a lot of this work quietly.
- Third-party mentions beat everything on your own domain. Budget real time for it.
- The pages are the easy part now. Mine went live in 35 minutes. Deciding what to build is the job.
The best use I have found for an AI coworker is not typing faster. It is having something that will argue with a bad plan at 9:45 on a Saturday and then execute the better one before lunch.
More on the AEO side of this: how I set up llms.txt so AI tools can actually read my site. And the long version of what it is like working with this thing daily is in my full Viktor review.
The tool
The AI coworker in these screenshots is Viktor. It lives in my Slack, has access to the sites and accounts I gave it, and does the work in the thread instead of handing me a to-do list.
If you want to try it, my referral link gets you $100 in free credits plus $50 off your first purchase, $150 total value. No credit card required to start.
FAQ
Do niche landing pages help you get recommended by ChatGPT?
Relevant, specific content does help, because AI answer engines run a live web search and pick the most contextually matched result. Thin templated pages do not help, because they tend to get suppressed in search and never make it into the model's source set.
How many industry pages should an agency build?
5 to 8 real verticals, each with content that is actually different. Mention sub-niches inside those pages rather than giving every sub-niche its own page.
What matters most for AI recommendations?
Third-party mentions. Listicles, review sites like Clutch and G2, industry roundups, and case studies other sites link to. Models weight what others say about you far above what you say about yourself.
How long did the build take?
About 35 minutes for 6 pages plus a new nav dropdown, then roughly 20 more minutes of my corrections on mobile styling and one real bug.
