80 impressions. Our best-performing blog page in the last 28 days — by far. I pulled the query breakdown expecting to see what people were actually searching for when they found it.
Every visible query contained the name of a company that does not exist.
The setup
Blog generation on valuefy.app is automated. A scheduled routine picks a topic, hits a Supabase edge function, and the edge function calls Gemini with a large system prompt assembled from prompt-builder.ts. That prompt contains, among other things, a block of fictional case study companies — examples injected so the AI produces naturalistic-sounding content instead of abstract theory.
One of those companies is TechFlow Solutions: an IT consulting firm in Stockholm, revenue SEK 22M, shifted from services to SaaS, hybrid model confused buyers, eventually split the business and sold the services side at 4x.
TechFlow Solutions does not exist. It was written into the prompt template so the AI would write concrete stories instead of generic advice.
The blog post "The valuation reset: navigating lower exit multiples in today's M&A market" was generated using that prompt. The AI apparently used TechFlow Solutions as an example. Google read the post, indexed the company name alongside the M&A due diligence language surrounding it, and decided: this is the right page to show when someone searches for TechFlow Solutions' financial data.
What I expected vs. what the data showed
The valuation reset post sits at average position 7.1 with 80 impressions over 28 days. That's the most visible blog page in GSC right now. I expected the query breakdown to show something like "exit multiples 2024" or "lower valuation M&A" — the actual topic of the post.
Here's what came back:
| Query | Impressions | Position |
|---|---|---|
| "techflow solutions" "normalized ltm ebitda" or "qoe" or "quality of earnings" | 6 | 7.0 |
| "techflow solutions" "qoe" or "quality of earnings" or "normalized ltm ebitda" | 5 | 7.0 |
Two rows. Eleven impressions total. Out of 80.
The other 69 impressions don't appear in the query breakdown — they're below GSC's privacy threshold, one or two impressions each, invisible. But based on the two that surface, the pattern is clear: someone is searching for "TechFlow Solutions" alongside specific M&A due diligence terms. That's not a person looking to understand exit multiples. That's due diligence research on a specific company — combined with the kind of financial language a quality-of-earnings analyst would use.
A company that was invented inside a TypeScript file.
Finding #1: Fictional names in prompts become real search keywords
The chain, traced through the code:
-
supabase/functions/_shared/prompt-builder.tscontains a fictional case studies block - TechFlow Solutions is one of five fictional companies at line 157: Stockholm, IT consulting, SEK 22M revenue, hybrid model problem
- The blog generation prompt injects this block for every valuation-type content request
- The AI used "TechFlow Solutions" in the generated post alongside real M&A vocabulary — "normalized LTM EBITDA," "QoE," "quality of earnings"
- Google indexed those phrases together
- The post now ranks at position 7 for search queries that combine the fictional company name with professional due diligence terminology
The impressions are not from people interested in exit multiples. They're from people — or more likely automated research tools — that encountered "TechFlow Solutions" somewhere and went looking for more. Position 7 for a fictional company is not a success. It's 80 impressions from completely the wrong audience, plus zero clicks because the page can't give those searchers what they actually want.
Finding #2: The advertising cluster follows the familiar pattern
This week the automation improved seven more calculator pages: Valuation Multiple (June 21), CPC (June 22), CTR (June 23), ROAS (June 24), CPA (June 25), Ad Spend (June 26), Burn Rate (June 27). Five of the seven are marketing calculators — the advertising cluster.
The CTR calculator is the only one with any GSC presence in the 28-day window. The page data shows 19 impressions, position 5.5, one click, and a 5.26% CTR. Position 5.5 on a competitive marketing term. That looks promising.
Then the query drill came back with zero rows.
Not low positions — no visible queries at all. All 19 impressions are below GSC's threshold: dozens of micro-impressions from long-tail variants, none large enough to appear. The average position 5.5 is arithmetic across an invisible distribution, the same pattern that's appeared for five consecutive clusters now. The other five advertising calculators don't appear in the top 50 pages even in the full 28-day window.
One note on the CTR calculator specifically: it does have a decent title from the automation — "CTR Calculator: Click-Through Rate & Benchmarks | Valuefy" — and a meta description that mentions specific platform benchmarks. The content improvement is real. The traffic just isn't there yet, and may not be given the competition for those terms.
Finding #3: 42 days, one complete cycle, now on the second pass
The daily automation has run without interruption since May 16: 42 consecutive commits, one per day, one calculator page per commit. The commit on June 27 was the Burn Rate calculator — same page as the very first commit on May 16. The queue has cycled back to the beginning.
That means we now have one complete pass through 42 pages, and the second pass has started. Whether improving a page for a second time — now that it's been indexed and potentially settled into a position — produces different results from the first pass is the question I don't have data to answer yet. I'll check back once the second pass is a few weeks in.
What I'm going to do about it
-
Replace named fictional companies with anonymous descriptions — "a Scandinavian IT consultancy with SEK 22M revenue" creates no searchable artifact. "TechFlow Solutions" does. The fix is to remove the names and let the financial details carry the example. Audit all five fictional companies in
prompt-builder.tsbefore the next content generation run. - Check whether other blog posts rank for the other fictional company names — TechFlow Solutions is one of five. Brightside Care, TechCo, and two others are also in the prompt. If other posts are ranking for "Brightside Care" plus healthcare M&A terms, this is a site-wide pattern, not a one-post anomaly.
- Hold judgment on the second automation pass — the useful question now isn't "did improving 42 calculators work?" It's "does a second improvement on an already-indexed page move positions?" That data arrives in a few weeks. I'll report what it shows.
The uncomfortable lesson
The fictional company names were put into the prompt to make the output sound better. Not to deceive anyone — there's no intent to mislead — but because "a company called TechFlow Solutions solved this exact problem" reads more naturally than "imagine a company that solved this exact problem." Concrete examples. Specific details. The usual advice for good writing.
The problem is that "better-sounding content" and "well-indexed content" pull in different directions when the named examples are invented. Content that sounds naturalistic to a human reader creates exact-match fingerprints in the search index. Google doesn't know TechFlow Solutions was typed into a prompt last year and that no actual business by that name exists. It sees a document containing "TechFlow Solutions" and "normalized LTM EBITDA" together and decides: when someone searches for both, show this page.
Named fictional entities in AI prompts are an SEO liability, not just a content-quality concern. Every invented company name, product name, or made-up case study is a potential keyword anchor that might attract queries for the wrong audience. For a valuation blog post trying to rank for "exit multiple trends," phantom impressions from TechFlow Solutions research queries are noise at best and a ranking dilution signal at worst.
The fix is narrow: strip the names, keep the numbers. The lesson is broader: when you're writing prompts for content that will be indexed, every named entity is a keyword decision — including the ones you invented.
I'll pull the query data again in four weeks. If the renamed prompt clears out the phantom impressions, that's a data point worth publishing. If the page is already permanently associated with TechFlow Solutions in Google's index and the queries persist, I'll say that too.
I'm running these experiments on valuefy.app and writing up what actually happens, including the parts that go sideways. If you're building with AI-generated content, debugging unexpected impressions, or working through the same "content looks right but rankings don't follow" problem, I'd love to compare notes — drop a comment or reach out.
I also run AImiten, where we build AI tooling for companies. This side project is where I stress-test ideas before they touch client work.
Top comments (0)