Our first 8 blog posts earned 86 Google impressions and 1 click in 3 months. They were well researched, sourced and carefully edited. The problem was upstream: we picked topics we found interesting and never checked whether anyone searched for them. Now we use an AI agent for SEO in a different order. It checks search demand first, matches that demand to something we actually built or broke, drafts only then, and nothing goes live until a person approves it.
This post is the exact process, with our real numbers, including the ones that are still too early to judge.
What Search Console showed
Between 1 August and 1 November 2026, venturehand.com had 1,263 impressions and 31 clicks in Google Search. 22 of those clicks went to the homepage, mostly people searching for our name.
The 8 blog posts we published before October:
- 86 impressions between them
- 1 click
- the best one, on the backlash against AI slop, had 35 impressions at an average position of 13
For comparison, our free tool and playbook pages did better with less writing. The /playbook/indie-hackers page alone had 185 impressions, even at an average position of 46.
Those posts were not bad. They answered questions nobody was typing. If you are writing blog posts for a new SaaS and your Search Console looks like this, the writing is probably not the first thing to fix.
The topic that made us change the method
In late October we had a post drafted about getting the LinkedIn Company Page API to work. It was first-hand, specific and useful. Aleksandar pushed back that nobody searches for it, so we had the agent check the volume for what a founder would actually type. "linkedin api post to company page" came back at about 10 searches a month in the US.
That is when we stopped choosing topics by what we found interesting. The post is still a draft.
The 5-step process we use now
The agent is an AI model connected to 3 things: our Search Console data, a keyword research tool for search volume and difficulty, and our own build logs. Here is what it does on every run, in order.
1. Read what already ranks. It pulls Search Console first: which pages get impressions, for which queries, at what position. Pages at positions 5 to 20 with real impressions are the cheapest wins, because Google already considers them relevant.
2. Turn our unused stories into searches. We keep a log of everything we built, measured, broke and fixed. The agent rewrites each unused story as what a founder would type into Google when they hit the same problem. Not "verification that reports success", but "why are my emails going to spam".
3. Check demand before writing a word. Each keyword group gets real volume and difficulty. Our rule of thumb: a topic goes ahead only if its keyword group gets 200 or more searches a month at difficulty 30 or under, or if something about it changed in the last 14 days. For this post, "ai seo agent" showed 260 searches a month at difficulty 17, and "how to use ai for seo" showed 140 at 28.
4. Drop anything a model could write without us. If a competent model could give the lesson with no access to our logs, the topic is out, however popular it is. Search demand decides what readers want. Our own material decides what we can say. A post needs both.
5. Draft against a hard floor, then wait for approval. Our blog connector refuses any draft under 700 words, without at least 3 FAQ entries, without an internal link, or with a statistic that has no source URL. It tells the agent every reason at once. A draft that passes is still only a draft: it sits behind an approval link, and Aleksandar approves or rejects it. The agent cannot approve anything.
What the first post from this method found
The first topic the new method picked was our cold email failure. The keyword group, led by "why are my emails going to spam" and "dmarc fail" at about 590 searches a month each, added up to roughly 2,000 searches a month at difficulty 3 to 18. Our own material was a mail-tester score of 1.5 out of 10 and the DMARC failure behind it. That became our post on why emails go to spam, published on 29 October.
It has no Search Console data yet. As of 1 November it had not shown up in impressions at all, which is normal for a 3-day-old page on a new domain. So here is the test we are setting for it:
- Pass: by 1 December, the posts chosen this way have more impressions between them than the first 8 posts earned in 3 months.
- Fail: they do not. Then search demand is not the bottleneck for us, and the next thing to look at is links and crawling, not topics.
What the agent did not catch
Demand-first topic selection does not fix technical problems, and ours had 2 that no amount of good topics would have solved.
A stale sitemap meant Google was reading a version that did not include 5 new posts. Separately, a navbar row we deleted had been the only internal link to 6 pages, so Google discovered them and never crawled them. Both were found by reading the Page Indexing report and our own code, not by the agent's topic research.
If you are starting out, check those basics before you write more. Our free SEO page analysis checks one page's title, description and other on-page basics and tells you what to fix.
How to try this on your own blog
You do not need our setup. With Search Console and any keyword tool, the same steps work by hand:
- Export 3 months of Search Console data by page. Note which posts have impressions and which have none.
- List the problems you have solved while building. For each, write the exact phrase you typed into Google when you hit it.
- Check volume and difficulty for those phrases. Keep groups with real searches and low difficulty.
- Cross out anything a generic article could cover just as well. Keep what only you can write.
- Set a pass/fail date before you publish, so you learn whether the method worked instead of guessing.
An AI agent makes steps 1 to 4 fast. It does not replace step 5, and it should not publish anything you have not read. We wrote about that side of it in why an agent should ask before it ships.