Hear Homer: the quote was sourced. It was about the wrong book.
For readers choosing which Odyssey audiobook to start with.
Hear Homer helps people pick an Odyssey audiobook, and I checked every spec on it by hand. The worst errors were in the opinions, so the build now asks every quote what it is evidence for, and refuses to run when one can't answer.
- 3 daysfrom the first commit to the first Copilot citation, July 21, 2026
- 346Copilot citations from Jul 20 to Oct 11, 2026, across 11 pages. Small numbers; the speed is the point
- 15 of 15planted defects the opinion gate caught, the empty-file case included
The whole site, month by month
- September vs August-11%8 search clicks, from 9
- Best month on recordAugust9 search clicks
Search clicks down 11% in September, after a record August.
| Month | Search clicks | Change from the month before |
|---|---|---|
| August | 9 | first month |
| September | 8 | -11% |
Two choices at once, and a test of my own method
Someone walks out of the 2026 film of The Odyssey, or gets the poem assigned, and wants the real thing in their ears tonight. Every recording is two choices at once: a translator and a narrator. The facts that settle it, like the runtime, abridged or not, and who reads it, were scattered across store listings and old press roundups. A site about which translation to read already existed. Nobody covered which recording to listen to.
I built for that person. I also wanted an answer to a question about my own method: could a brand-new domain with no authority earn AI citations by being the most careful source on a narrow question? The day after launch an audit kept pointing at zero citations. The point was to read what the AI answers cited that day and work out whether this site would eventually be one of them.
A registry for the specs, then an argument for the opinions
19 URLs in the sitemap, one registry file for every spec. A spec appears on a page only after I've matched it to the recording's own listing, and each page shows the date its specs were last verified. No prices, because they change every day. Real search queries decided the questions each page answers.
Then I noticed the imbalance. I had checked runtimes obsessively and reviews only casually, and nobody picks a recording by its runtime. So I published an argument for and against each lead recording, quoting critics and listeners, and gave every outside quote a rule: it needs a link, a publication date, a last-verified date, and what the quote is evidence for: this recording, this translation, or other work. The build refuses to run when a quote is missing any of those.
The pages people reach most:
Sitemap counted on the live site, Oct 13, 2026. The quote rule is quoted from the site's build script.
Cited in three days, then a review pass found four live errors
Bing's Copilot cited the site 3 days after the first commit, on July 21, 2026. By Oct 11, 2026 it had 346 citations across 11 pages, and the page cited most was /every-odyssey-audiobook/, with 103. For the query "best odyssey audiobook", this site was the source in 67% of the answers that cited anyone.
The review pass behind the opinion gate found four live errors in July 2026. The movie page said the film leaned on one particular translation; that traced to a single magazine mention the magazine had since walked back. One page quoted reviews of a different book. Awards were stated as fact where the registry said "publisher claims". A fifteen-minute video was called a short preview. Every one was sourced. Every one was wrong.
In launch week, Google sent 117 clicks and GA4 recorded 20 visits. I checked one tool against the other, found a security setting silently throwing GA4's data away, and fixed it before anyone made a decision on bad numbers.
Google is the honest miss. After that first week, Google sent 0 clicks in August 2026 and 2 in September 2026. Bing impressions went from 127 in August 2026 to 159 in September 2026. I have not diagnosed the Google drop, and I'm not going to guess at a cause here.
How I measured it. Citations and the query: Bing Webmaster Tools' AI Performance export, a manual pull that stops at Oct 11, 2026. Launch week: Google Search Console against GA4, data through Jul 24, 2026. Clicks and impressions by complete month: Search Console and Bing Webmaster Tools, through Oct 10, 2026. All small numbers; read them as direction, not volume.
A real, dated, linked review can still be about a different book
My name went on an anonymous site
On July 19, 2026 a trust pass added a byline with my name and city to a site I had built to stand on its sources, not on me. I had it removed the same day.
The gate trusts what a quote says about itself
The rule checks that every quote declares what it is evidence for. It does not check whether the declaration is true. The McKellen page dates its recording to 1999, and two of the reviews it quotes as evidence about that recording are from January 1997 and July 1997. A review can't be about a recording released after it. Either the date or the tag is wrong, and I haven't settled which release those critics heard. It's live as I write this.
What changed, and what's next
Checking that a source exists wasn't enough. A review can be real, dated, linked, and about a different book. So every quote now says what it is evidence for, and the build won't run when one doesn't. I proved that by planting 15 defects, the empty file included, and watching all 15 fail. The next step follows from the McKellen miss: the quote's date has to agree with the recording's date, and a machine has to compare them, because a field that states the right answer isn't a check.
The byline and the gate dates are from the site's commit history; the McKellen dates are from its data files and the live page, read Oct 13, 2026.
The evidence
