Cyrus Shepard Found the 4-Trait Cliff: Site Win Rates Jump From 30% to 68%
Cyrus Shepard analyzed 400+ websites for his Zyppy Signal newsletter on April 9, 2026 and found a sharp inflection in 2026 Google traffic outcomes. Sites with three of his five Google-friendly traits won the algorithm 30.7% of the time. Sites with four traits won 68.1%, more than doubling the odds and isolating the only threshold that matters for SEO planning this year.
The full study covers a 12-month traffic window, revisits the cohort from Lily Ray's December 2025 update analysis, and uses Spearman correlation to score five features against year-over-year traffic change. PPC Land's writeup is a useful summary if you want the bullet version. I've read both twice. The headline isn't the five traits, which most SEOs will recognize. The headline is the cliff between three and four. Everything below is built around that.
The five traits, in plain English
Shepard ranked the features by how strongly each one correlated with traffic gains. The order matters because it tells you where to spend planning energy if you can only fix one thing.
Offers a product or service (0.391 correlation). 70.2% of winning sites sold something they made themselves, against 34.6% of losers. The winners include budgetbytes.com and mathnasium.com. The losers include byrdie.com and medicalnewstoday.com. Subscriptions and digital goods count.
Allows task completion (0.381). 83.7% of winners let the user finish the job they searched for, on the page they landed on. Buying, downloading, calculating, booking, searching a database. 50.2% of losers can do this. The illustrative loser in Shepard's writeup is WalletHub: a credit card comparison page that doesn't let you actually apply for the card.
Owns proprietary assets (0.357). 92.9% of winners own something other sites can't easily clone. A dataset, a community, a piece of software, a UGC corpus, a review archive. 57.1% of losers do. Letterboxd and Todaytix sit on the winning side. Lifewire and The Spruce sit on the losing side.
Tight topical focus (0.250). 75.9% of winners cover one narrow thing in depth. Compare minecraft.wiki to businessinsider.com and the difference becomes obvious. Broad publisher coverage seems to be a structural liability now, not a strength.
Strong brand search (0.206). 32.6% of winners pull a meaningful share of branded navigational queries. 16.1% of losers do. zoom.com and skims.com show up here. The bar is lower than people assume; you don't need to be Coca-Cola, you need to be the thing people type into Google by name.
Why the jump from three to four is the only number worth your planning time
Here is the additive table from Shepard's analysis, win rate by feature count:
0 features: 13.5%
1 feature: 15.4%
2 features: 22.0%
3 features: 30.7%
4 features: 68.1%
5 features: 69.7%
From zero to three, every added feature buys you roughly 5 to 9 percentage points. Worth doing, not transformative. Then the fourth feature adds 37.4 points in a single step. The fifth adds 1.6.
That shape is unusual. Most SEO factors I've worked with show either linear lift or diminishing returns, not a cliff. A cliff means there is some interaction effect Google's systems are reading once a site crosses an internal threshold of "real business that does a thing." Three features reads as content. Four features reads as utility. The algorithm seems to weight that distinction harder than it weights any individual signal. ALM Corp's interpretation goes further and frames it as a brand-demand prerequisite for any AI Overview citation; I think that's plausible but oversold on the available data.
From what I've seen, this also means most agency proposals will sell you the wrong thing. Pitching a client on "cover all five" is overengineering. The cheaper, faster move is identifying which one missing feature is closest to existing capability, and shipping just that one. For a content-heavy publisher with strong brand and topical focus, that's almost always task completion. Build a calculator, a comparison tool, a database lookup. Anything where the user can do the thing instead of just reading about it.
What "winning" looked like, in named examples
Shepard's named winners read like a who's who of utility-shaped sites: stockanalysis.com, mathisfun.com, powerball.com, happiestbaby.com, Letterboxd, Todaytix. None of them are content brands. They're databases, calculators, communities, ticket engines, lookup tools. Even mathnasium.com, which looks like a content site, is really a tutoring booking funnel.
The losers read like an obituary for the 2010s SEO playbook. Byrdie. Medical News Today. Lifewire. The Spruce. TechTarget. Newsweek. Business Insider. These are sites that built their traffic on long-tail informational queries that Google now serves with AI Overviews and a citation. We covered the same dynamic from Google's own side last week when Danny Sullivan named "commodity content" as the SEO liability killing blogs. Shepard's data is the receipts.
The pattern isn't about content quality. Byrdie is well-edited. Lifewire has named subject experts. Medical News Today has actual medical reviewers. None of that is enough anymore. The question Google appears to be answering is closer to: does this site exist for a reason other than ranking? If the answer is no, the cliff catches it.
What didn't correlate, and why that should make you nervous
Shepard tested several features that did not show a meaningful correlation with traffic gains. First-hand experience signals. UGC volume. Information uniqueness. The full E-E-A-T canon, basically.
His read is that these are now baked into the algorithm at a baseline level, so they're table stakes rather than differentiators. That's the charitable interpretation. A less charitable read is that publishers who spent 2024 and 2025 retrofitting bylines, expert reviewers, and original photography are no longer getting credit for the upgrade because everyone else did the same thing. The arms race ended with everyone in roughly the same trench.
I think the operational implication is uncomfortable. If E-E-A-T work doesn't move the needle in a measurable way, but proprietary assets and task completion do, content teams have been pointed at the wrong scoreboard for two years. Editors and standards are still important for not getting demoted on a future helpful-content style update. They're just no longer a path to growth on their own. We saw a related signal earlier this month when Liz Reid conceded AI Mode queries run 2-3x longer than keyword tools track, which fits the same theme: content that answers a single query is increasingly being intercepted before it gets a click.
One prediction with a number attached
I'd put the over-under at three: I think we see at least three top-100 US publishers (think About.com-era brands, the ones currently on the loser list) acquired, downsized, or shuttered in the next 18 months as the 30/70 cliff becomes a CFO conversation rather than an SEO one. The math is brutal. A site with two of Shepard's features doing 50M monthly sessions in 2024 is, on average, doing 22% of its old traffic by the end of 2026. That's not an SEO problem you optimize out of. That's a business model.
And honestly, the more I sit with this, the more it reads like Google deciding what kind of internet it wants. Less indexable, more transactable. Whether that's good for the open web is a different essay.
A 30-day audit worth running before next quarter's plan
Take your top 50 indexed pages by impression. Score each one against Shepard's five features as the page itself, not the site. Most content pages will land at one or two. If your site as a whole is sitting at three features and your individual pages are sitting at one, the gap is your action item. The page is what Google ranks. The site is what Google trusts.
For most blog-shaped sites I've talked to in the last year, the cheapest fourth feature to add is task completion. Pick the five highest-impression pages without a tool on them. Add one. A calculator, a checker, a comparison engine, a generator, anything that takes input and produces output. Ship it within the quarter. Then look at the impressions-to-clicks ratio 60 days later and see if it moved.
I don't think Shepard's framework is the final word. The correlations aren't causation, the sample is biased toward sites Lily Ray was already tracking, and 12 months is short for an algorithm that updates monthly. But the cliff between three and four traits is too sharp to be noise, and the cost of acting on it is low. If it works, you've moved a structural lever. If it doesn't, you've shipped a useful tool. That's a much better trade than another 6,000-word commodity article.
Notice Me Senpai Editorial
