Google Discover Data: What I Learned From 16 Months of Traffic

Reading time: 11 min

Google Discover Is Not a Bonus Channel Anymore

I’ve been running websites longer than most people have been online. But Google Discover still caught me off guard. For years, I treated it as nice-to-have traffic—welcome when it arrived, baffling when it vanished. Not anymore.

On Minecraft.fr, Discover now drives over a million visits a year. On good months, it accounts for a third of all SEO traffic. That’s not a bonus—it’s a core channel. So I spent four months digging into 16 months of Google Search Console data, a full crawl of Discover pages, WordPress export of every article, and four months of ad platform data.

Here’s what actually happened.

Search and Discover Tell Different Stories

Open data across both channels simultaneously. The contrast is stark.

Search traffic contracted gradually from summer 2025 onward. Discover traffic exploded in late 2025 and kept climbing into early 2026. The editorial spikes in Discover are no longer accidents—they’re filling the gap Search is leaving.

Comparing January–April 2025 to January–April 2026, the numbers are clear. Search lost 38% of clicks. Discover gained 67%. Combined total is still negative (roughly -28%), but the composition of SEO changed entirely.

In January 2025, Discover represented 10% of monthly SEO traffic. By January 2026, it was 33%. Tripled in twelve months.

Let me show you the data.

The 72-Hour Sprint: Discover’s Peak and Decay

Look at individual page life cycles. Over 198 articles that generated at least 500 Discover clicks, the pattern is remarkably consistent.

A page hits its peak at Day+1 in most cases. At that peak, it does a median of 37% of all its future Discover traffic. Publication day itself? Only 10%—Google needs time to find and push it into feeds. Then it’s a bell curve: Day+2 at 19%, Day+3 at 9%, Day+4 at 5%. By Day+7, nearly zero.

Superimpose the trajectories of the top 10 pages—they all converge on the same shape. Some peak at Day+2 instead of Day+1, amplitude varies, but the form is identical.

Key statistic: 66% of pages hit their peak on Day+1, 94% by Day+3 at the latest. Total exposure window hovers around a week and a half. Very few last beyond two weeks. Exceptions are seasonal revivals or rebounding topics.

Here’s the counterintuitive part: the correlation between page lifetime and total clicks is 0.16. Virtually zero. A page that spikes hard on Day+1 will generate more total clicks than a page that trickles along modestly for a week.

This changes how you think about Discover. It’s not a continuous feed channel with steady returns. It’s a peak-at-72-hours mechanism. You must win on Day One.

Zero Overlap: Search and Discover Reward Different Pages

I expected some overlap between Search and Discover top pages. The data was more extreme than I imagined.

Over 16 months, the top 50 Discover pages and the top 50 Search pages share zero URLs. Zero. Top 100 vs. top 100? Exactly one. For URLs active in both channels (at least 100 clicks each), the correlation between Search and Discover volumes is -0.04—statistically zero.

Concretely: the pages that get the most Discover clicks also appear in Search, but at an average position near the bottom of page one (around 7th spot) with Search traffic that’s dozens of times lower than their Discover volume.

Discover pages are fresh news: update tests, mod announcements, collaborations—topics with a short relevance window. Search pages are the opposite: evergreen guides, technical FAQs, mod files, category pages. Two types of content, two visibility logics, and essentially no overlap at the top.

One nuance: a page that breaks on Discover is inherently recent—it hasn’t had time to accumulate Search equity like evergreen guides. But the absence of overlap is real, driven by algorithmic differences, not just age.

Nobody talks about this part: optimizing for Search does not help Discover. They’re separate games.

What Makes a Discover Page Perform: Signals from the Data

I crawled every URL that appeared in Discover over 16 months to find common patterns. Here’s what jumped out.

Titles with a question mark show a weighted CTR of 6.49%, versus 4.39% average (+48%). List-style formats (“Top X”) hit 5.91% (+35%). Only 16 out of 238 pages use these formats—massively underutilized leverage.

Mentioning a year or version (1.21.9, 26.1) doesn’t affect CTR, but pages with these mentions get two to three times higher median clicks. They capture larger audiences.

Article length: 1,200 to 1,800 words get five times the median clicks of 500-800 word pieces. CTR barely moves (4.3% to 4.7%), but total volume captured increases with length up to about 1,500 words, then plateaus. Discover doesn’t reward depth, but longer articles have more chances to be recommended to a wider audience.

Category differences are massive. Using “Minecraft News” as baseline (index = 100), all calendar-linked categories outperform: “Update” at 249, “Minecraft Live” at 197, “Minecraft 1.21” at 170. Evergreen categories (Mods, Modpacks, Resource Pack, Shaders) range from 35 to 60. Step away from immediate game news, and Discover performance collapses.

Image count: at equal length (800–1,500 words), jumping from 6–15 images to 25+ images multiplies median clicks by 1.7. More visuals give Google more choices for the feed card image.

Opg:title vs. H1: on 267 crawled pages, 122 (46%) have a different og:title than the H1. Across all categories, the effect is marginal (+21%). But isolated to Minecraft News (80 pages) and stratified by length (800-1,500 words), the net effect is +48%. The raw view shows +77%, but that’s partly because pages with custom og:titles are longer on average—the +48% is the clean number.

This matters because WordPress H1 is constrained by SEO best practices. But Discover displays the og:title, not the H1. Rewriting it to be more journalistic—like a press hook—increases card CTR. A lever many overlook because it requires manual per-article intervention and doesn’t appear in standard SEO tools.

Three Real-World Og:Title Customizations

These examples show the “how,” not the “how much.” The isolated title effect remains the +48% measured above.

Case 1 – Minecraft News: ×11 vs. site median

H1: “10 improbable Minecraft collaborations we dream of seeing”
og:title: “Minecraft x Tamagotchi, Durex, IKEA… when blocks invade real life”
Generic listicle replaced with concrete teaser using iconic brands.

Case 2 – Minecraft News: ×7.6 vs. site median (recent test)

H1: “Minecraft Java finally gets an in-game friends list”
og:title: “Mojang finally adds this feature to Minecraft after fifteen years of waiting”
H1 describes the feature; og:title hides it, creating emotional anchor (“finally,” “fifteen years”) and curiosity.

Case 3 – Modpacks: ×7 vs. site median

H1: “BigChadGuys Plus, the Minecraft modpack that blends relaxation and Cobblemon”
og:title: “This Minecraft modpack combines Stardew Valley and Pokémon in a peaceful world”
Obscure name removed; niche reference replaced with universal pop culture (Pokémon, Stardew Valley).

Two strategies emerge: either remove the obscure proper noun for a pop culture reference, or remove the precise description for emotional anchoring. H1 describes the page, og:title sells the click.

Three Custom Og:Image Experiments

Fewer pages (22 out of 267, or 8%) have custom og:images. Sample too small for statistical significance, but the “how” is instructional.

Case 1 – Minecraft News: #1 all-time Discover hit (×41 vs. site median)

Post thumbnail: a composed scene of ten baby animals (cluttered subject).
og:image: side-by-side comparison of two chicks with dates “2011” and “2026”—meme format “before/after.”
Readable in under a second. Also reused from an already-viral meme on X and Reddit—

But recycled with permission and proper format by me. Recognition effect in feed.

Case 2 – Minecraft News: ×7.6 (deliberately incongruent image)

Post thumbnail: split-screen literal screenshot of the Friends menu + Alex/Steve high-five.
og:image: Minecraft character alone, mouth wide open in a “mind blown” meme style. No visual link to the feature. Test: the image doesn’t show the topic, it stops the scroll by incongruity. I also pushed the og:title hard: “Mojang finally adds this feature to Minecraft after fifteen years of waiting.”

Case 3 – Resource Pack: ×11 vs. site median

Post thumbnail: character in wise posture with overlay “ENDERWOMAN – TEXTURE PACK.”
og:image: feminine, stylized Enderman silhouette, no text. Provocative, borderline for Discover. It passed. Lever probably the removal of branding and the visual appeal.

I don’t have a formula for og:image yet. I’m testing. All three images were AI-generated with Gemini—no issues from Discover.

The Weekend Effect: When to Publish for Discover

Sunday is the best day for Discover clicks—+41% versus Tuesday, the worst. For a leisure site like Minecraft.fr, this makes sense: Sundays mean more time, phones in hand, scrolling Google feed.

Combine this with the Day+1 peak: aim to publish late in the week so the peak lands Saturday-Sunday.

Publication Cadence: Discover Is Probabilistic

Cross-referenced article export with daily Discover traffic. The first finding isn’t flattering.

Not one day of visibility for most articles. They’re published, indexed, but the channel doesn’t push them. At the other extreme, about 5% become hits, and 23% generate solid volumes. The middle third performs modestly. The distribution is extremely unequal: most posts vanish, a minority capture nearly all traffic.

The hit rate varies by quarter—between 38% and 61%—without a clear editorial pattern to explain it. When Minecraft news is rich, Discover pushes more. When it’s flat, the channel stays flat regardless.

Publishing multiple articles in one day? At close approximation, no effect on individual article performance. A single article has the same profile as groups of two or three. Groups of three have a slightly later peak (Day+2) and longer tail when they cover the same topic. Marginal effect.

The inverse effect is brutal. I left for two weeks in April 2026—13 days without publication. Discover traffic faded over 3-4 days as last articles cycled out, then collapsed: from 128 clicks to zero within seven days. The channel doesn’t pause politely—it turns off in under ten days.

At monthly scale, the correlation between publication volume and Discover traffic is 0.60 (Spearman). Solid, but caution: on update-heavy months, I publish more AND the channel is more generous—hard to isolate cause and effect.

Bottom line: more publications don’t guarantee more Discover. January 2026 (34 articles) outperformed April 2025 (44 articles). Content, timing, and luck matter as much as frequency.

Follow Someone Else’s Calendar to Predict Discover

The channel is probabilistic, but the probability depends on something publisher-controlled: timing to external events. I tested a linear regression on content variables (length, images, category) to predict per-page Discover clicks. R²? 1.7%. Everything measurable about the page explains almost nothing. The other 98% depends on external factors.

Cross-reference Discover spikes with Minecraft’s official event calendar (Minecraft.net, wiki). Result: of the top 30 Discover traffic days over 16 months, 24 coincide with a major event—Minecraft LIVE, game drop, snapshot covered by gaming press, or transmedia event like the film release (April 2025).

The absolute peak, January 9, 2026, wasn’t a release day. It was two days after snapshot 26.1, when Windows Central, PC Gamer, Beebom, and all gaming press simultaneously published about the new baby mobs. Pure media buzz, no release—and it was the window where any Minecraft.fr article on that topic had a shot at Discover propulsion.

Traffic multiplier during 14-day windows around major events: 4x to 10x compared to quiet periods. Conversely, Mojang’s editorial lulls (July–August 2025, February 2026) stay at a low floor regardless of what you publish.

Think this through: if you map exogenous event calendars—keynotes, product launches, sports seasons, game drops, movie release dates—you map exactly when Discover becomes somewhat controllable.

Recovery Is Asymmetric: Off in 7 Days, On in 24 Hours

After the silence ended, I resumed near-daily publishing for two weeks. What happened next is telling.

From April 22–27, the channel decayed from 128 clicks to 5. From April 28 to May 1st: absolute zero (0 or 1 click/day). Seven days from normal to dead.

Restart clicks came the same day I published again. Day one: 33 clicks. In 48 hours, back above baseline historical levels, and stayed there for the next two weeks.

No penalty for the two-week absence. No probation period to rebuild trust. As soon as a new article enters the evaluation cycle, the channel resumes as if nothing happened. Reactivation latency: hours, not days.

10% of Pages Drive Half Traffic

Discover’s per-page distribution follows a severe Pareto curve. Top 10% of URLs generate 48% of traffic. Top 20%: 66%. Top 50%: 92%. The other half: 8%. Gini coefficient 0.65.

Over 16 months, the top single article accounted for over 7% of all Discover traffic. Meanwhile, dozens of other published pages drive a fraction of that volume.

This confirms the probabilistic nature. One hit in a month counts for more than twenty mediocre posts. But the hit isn’t fully planable—it’s about topic, timing, luck. The only leverage: multiply the bets.

Discover Pays Better Per Slot Than Search

Now the unusual part. I crossed GSC with four months of ad platform data. To compare revenue value of a Discover click versus a Search click, I classified pages: “Pure Discover” (80%+ from Discover), “Pure Search” (80%+ from Search), “Mixed” (rest). Index with Pure Search as baseline = 100.

Result: at equal pageviews, Pure Discover pages earn 26% less than Pure Search. Sounds like Discover monetizes worse. But dig deeper: Pure Discover pages display a median of 2.5 times fewer ad blocks per session than Pure Search. Search pages are long, scroll-heavy, rich ad inventory. Discover news articles are short, read fast, limited space.

Normalize revenue per ad block displayed, and the picture flips. The Discover click is worth more, but over fewer units. The surprise is the Mixed cohort (15 pages, weaker statistically but notable): those pages earn +18% per pageview versus Pure Search. They combine decent ad density (evergreen life) plus high-value Discover visitors.

Implication: don’t bloat pages. But on high-potential Discover articles, add two or three extra ad blocks—there’s usually room without degrading UX.

After 16 Months, Here’s What I Know

Discover was a bonus channel for years. No longer. On a good month, it’s a third of total SEO traffic for Minecraft.fr. Stop observing it. Start treating it as its own channel with its own KPIs and cadence.

Five things the data forced me to accept:

• The window is 72 hours. By the time you notice an article spiked, it’s already delivered 80% of traffic.
• ~50% of articles never get any Discover traffic. The only countermeasure is volume—not for leverage, but for lottery tickets.
• Stop publishing, and the channel dies in 4–5 days. No memory on pause.
• Follow someone else’s calendar. Find the external event rhythm that drives your audience. Align publishing. Accept quiet periods as noise.
• Discover and Search are different games.

Also: the Discover click is more valuable than many think—not per session, but per ad slot. Revenue ceiling isn’t click quality, it’s ad inventory. Fix that operationally.

Slow down. Think. Then decide if you’re treating Discover like a bonus or a channel worth designing for.

This is a single-site analysis, gaming niche, non-optimal editorial frequency. The orders of magnitude (Day+1 peak, 72-hour window, zero Search overlap, ~50% hit rate) feel transposable to other editorial verticals. The absolute numbers are not. If you’re running a content site with Discover data, I’m curious what your numbers say. We’ll get a mature understanding of this channel only by crowdsourcing real data.

Two practical things I did while writing this piece: Since 2025, Google has generated a per-editor profile on profile.google.com. On June 4, 2026, Google formalized it as “Search profiles”—claimable, customizable, linked to Knowledge Graph, with beta Insights (Search Console data). U.S. only for now, but the auto-generated profile exists for non-U.S. editors identified as entities. I found mine via 1492.vision’s monitoring tool. If you dig into it,

The data submitted via this form is intended for Abondance only. It will not be transferred to third parties. You can unsubscribe at any time by clicking the unsubscribe links in each of our emails. For more information, please consult our full policy on the processing of your personal data.

The data submitted via this form is intended for the site owner. It will not be transferred to third parties. You can unsubscribe at any time.

Matt S. Riley

Building websites since before Google existed. I’ve run SEO, growth, and content for startups across California — and I’ve watched every ‘revolutionary’ tactic eventually expire. What doesn’t expire: understanding systems, compounding effort, and thinking slower than everyone else.