For UK tradespeople and local businesses publishing blog content, Google Discover is the wild card.
An article on the latest Google update or a guide on getting more reviews can sit dead for weeks, then suddenly land on the home feed of fifty thousand phones overnight.
The mechanism behind this has always felt random.
It is not.
Metehan Yesilyurt published research in March looking at the Discover Android app at the SDK level. The findings give the clearest picture yet of how Discover decides which articles to serve, which it ignores, and which it silently filters out before ranking even begins.
The Nine-Stage Pipeline
Discover runs articles through a structured nine-stage pipeline. The stages, in order, are:
Crawl. Read meta tags. Classify content type. Check publisher block status. Match to user interests. Apply server-side pCTR prediction. Build the feed layout. Deliver. Record feedback.
The order matters. Each stage filters out articles that fail it. Articles only reach the ranking step (the pCTR model) if they pass every prior stage.
The Publisher Block Happens Before Ranking
One of the most consequential findings: when a user taps "do not show content from this site", the block is applied at stage four, before content-to-user interest matching and before ranking.
The practical consequence is brutal.
If enough users block your domain, your articles never reach the ranking stage. You can have the best title, the strongest image, and the highest predicted CTR in the index, and none of it matters. The system filters you out before any of those signals are evaluated.
Crucially, there is no equivalent sitewide boost mechanism. Blocks compound. Endorsements do not.
What the App Sends Google Before Ranking
The SDK exposes which signals get transmitted to Google's servers before the ranking model runs. They are:
Your page title. Specifically the og:title meta tag, with fallback to the Twitter title tag, then the HTML title.
Your image. Including its dimensions and whether it actually loaded successfully.
Your content freshness. How recently the page was published and when it was last modified.
Your URL's historical click-through and impression data for users in similar interest groups.
The pCTR model itself runs server-side and is not visible to the app. But these are the inputs.
The Freshness Window
Discover groups content into time buckets, each receiving a different baseline boost:
One to seven days old. Strongest visibility. This is when articles peak in the feed.
Eight to fourteen days old. Moderate visibility. Significantly reduced from the peak.
Fifteen to thirty days. Limited visibility.
Thirty plus days. Gradual decline to near zero.
There is a separate classification for genuinely evergreen content that bypasses some of this decay, but the default behaviour heavily favours newer publishing dates.
Two implications for trade blogs. First, the recency boost is real and large, so seasonal articles matter, and publishing a fresh take on a topic competing publishers wrote eighteen months ago can win the feed slot. Second, articles older than thirty days have to earn their visibility through engagement signals rather than recency.
The Image Rules That Disqualify You Instantly
Discover reads six page-level tags. Two of them can stop your article appearing in Discover entirely.
The first is nopagereadaloud. The second is notranslate. If either is set, the article is excluded from Discover regardless of content quality.
The image rules are stricter still:
If your page has no og:image, Discover serves no card.
To qualify for the large prominent card format (the one that actually drives clicks), the image must be at least 1200 pixels wide.
Smaller images render as thumbnails and earn substantially fewer clicks.
For UK tradespeople, this single rule is the easiest fix and the most commonly overlooked. A 600 by 400 image is the difference between a quiet article and an article that brings in calls for a week.
Personalisation Is Layered
Four separate inputs combine into the personalised feed each user sees:
Google's broader interest data derived from cross-product behaviour. This includes search history, YouTube watch patterns, and Maps activity.
Publisher-level signals. Whether the publisher is registered in Publisher Center, whether they have been followed, and the publisher's reputation score.
Per-URL actions from the specific user. Saves, follows, dismissals.
Engagement signals. Time spent reading, scroll depth, returns to the article.
The dismissal signal is interesting because it is permanent. If a user dismisses a specific URL, that URL is stored and will not resurface in their feed.
Experiments Are Everywhere
During a single observed session, around 150 server-side experiments were running simultaneously. Another fifty plus feature controls affected how cards were displayed.
The practical consequence: two users with similar interests, in similar locations, looking at the same feed at the same moment, can see noticeably different content because they have been bucketed into different experiment cells.
This is why Discover traffic is so volatile. It is not just user behaviour driving variation. It is the system itself running constant A/B tests on every aspect of how content is selected and displayed.
What This Means for Trade Blog Publishers
Five clear implications for anyone publishing trade or local business content with hopes of Discover traffic:
One. Your og:image must be at least 1200 pixels wide. Without this, you do not get the large card and you do not get the click rate that comes with it.
Two. Publish freshness matters. Articles built on news hooks, seasonal topics, or recent Google updates will get a meaningful baseline boost compared to evergreen content.
Three. Once you irritate a user enough to block your domain, you are gone from their feed entirely. This is a real risk for publishers who reuse templates, ship thin content, or run intrusive ads. Trade blogs that look like content farms get blocked.
Four. Title quality is one of the small handful of signals transmitted before ranking. The og:title is doing real work. Generic templated titles compete poorly with titles that read like a clear answer to a question.
Five. The pCTR model rewards engagement. Time on page, scroll depth, and returns all feed back into the system. Articles that look like they will get clicked but bounce immediately get demoted quickly.
The Takeaway for UK Trades
For trade blogs, Discover is unpredictable but it is not random. The publishers who consistently win feed slots tend to share five traits:
They publish on news hooks within forty eight hours of the news breaking. They use og:images at 1200 pixels wide minimum. They write clean, specific titles. They keep their content templates varied enough to avoid the duplicate-content fold. And they have not accumulated enough dismissals to be filtered out of significant user populations.
If you are running a trade blog and not seeing Discover traffic, those five points are the order to fix them in.
We publish trade blog content that gets indexed and shared.
Every article includes a 1200x630 og:image, schema markup, and a unique content shape that avoids the duplicate-content fold. Included in every Local SEO Plan from £299/month.
📞 01902 961078 · Mon to Sat 8am to 6pm