Google Discover Mechanics: How Content Gets Chosen and Amplified

If one story surges in Google Discover while the next one disappears, it is tempting to blame timing, the headline, or luck. That diagnosis is usually too blunt. A page can miss the candidate pool, win attention but lose engagement, or satisfy readers yet reach too few people because the system has weak evidence that this audience and your publication belong together.

The useful shift is to treat Discover as a recommendation funnel with distinct jobs. Once you separate candidate retrieval, user-content prediction, final ranking, and learned affinity, you can identify the weak transition and work on the right problem.

Discover is a four-part recommendation system

Google groups Discover ranking work around retrieval, prediction, ranking, and embedding. These are not four optimization factors or a checklist for publishers. They are four technical jobs within a recommendation system:

1. Retrieval assembles a set of articles, videos, and other items that might suit the user.
2. Embeddings represent users and content in a form that allows the system to estimate similarity or relevance.
3. Prediction estimates what may happen if a particular card is shown to a particular user.
4. Ranking resolves the competing candidates into the feed the user actually receives.

The jobs interact rather than forming one simple, publicly documented sequence. Embeddings can support retrieval as well as prediction, and ranking can use information that publishers cannot observe. The model is still valuable because it stops you from treating every distribution problem as a headline problem.

Retrieval is especially easy to overlook. You cannot rank well inside a candidate set you never entered. Across 42 million monitored cards, about 20 candidate pipelines have been mapped, including candidate sampling, cluster-profile retrieval, trend-embedding retrieval, item-to-item collaborative filtering, and a post-retrieval pipeline heavily populated by YouTube and X content. The labels expose multiple routes into Discover, although they do not disclose the precise rule set behind each route.

A channel labeled as generative retrieval also appeared in September 2025 in roughly 0.03% of the French Discover feed. That tiny footprint is consistent with a limited test of model-driven candidate selection, not evidence that generative retrieval has replaced the broader system.

Observed user representations add another clue. Their names cover durable Discover interests, a short-term interest variant, trends, real-time behavior, and shopping-related behavior. This is consistent with a two-tower design in which user and content representations are compared in a shared vector space. The visible labels are real observations; the exact architecture and purpose of each representation remain interpretations rather than confirmed Google documentation.

Your practical response is to add an audience-state map to your keyword and topic planning. Before approving a Discover-oriented pitch, record:

• The intended reader: Name the person and existing interest the story serves. A broad demographic is less useful than a recognizable need or content habit.
• The time horizon: Decide whether the story serves an enduring interest, a developing trend, or an immediate event. Do not judge all three by the same distribution pattern.
• The relationship to previous coverage: Identify whether the story begins a subject, extends a cluster, or follows an item readers already encountered.
• The next useful item: Plan what a satisfied reader would reasonably want from your publication after finishing this page.

None of those fields forces retrieval. They make your publishing intent coherent enough to evaluate. If your team cannot explain who a story is for, why it matters at that moment, or how it relates to your established coverage, changing a few keywords is unlikely to solve the underlying recommendation mismatch.

Attention and deep engagement are separate predictions

Discover does not appear to reduce content quality to one universal score. About nine observed prediction values collapse into two nearly independent dimensions: whether a person is likely to stop on a card, and whether that particular person is likely to click and read deeply.

The correlation between those dimensions is close to zero. A card can be highly effective at interrupting the scroll while being a poor match for sustained reading. That is the mechanical form of clickbait: the promise wins attention, but the experience does not hold it.

The predictions also correspond with observed behavior. Interaction roughly doubled from the bottom to the top of the deep-engagement score range and declined as the predicted likelihood of scrolling past increased. These measurements do not reveal every ranking input, but they are strong enough to justify separating your own attention and engagement diagnostics.

Those publisher metrics are diagnostic proxies, not a list of disclosed Google ranking inputs. An increase in engaged time, for example, does not prove that one metric directly caused more Discover distribution. The purpose of the table is to locate the leak in your own experience before you prescribe a fix.

If impressions are meaningful but card response is weak, examine attention and candidate-to-reader fit. If clicks are healthy but readers leave quickly, the problem is downstream: the audience may be wrong, the opening may delay the payoff, or the content may not fulfill the card’s promise. If both look healthy but amplification remains limited, a more aggressive title is not the obvious next move. Retrieval, reader-source affinity, and usefulness still need investigation.

This distinction should change how you run headline tests. Evaluate the card response and the post-click session together. A variation that increases clicks while reducing reading depth may have widened the promise-content gap rather than improving the story’s overall Discover potential.

Reader-source affinity can outweigh topic potential

Topic relevance gets a page into the conversation, but personalization can determine how loudly it is heard. Reader-source affinity is the learned relationship between a specific person and a specific publisher. It is not identical to general popularity, topical relevance, or the number of people who pressed Follow.

A small comparison involving two French sports publishers with nearly equal topic potential illustrates the possible size of that effect. The publisher with deep-engagement predictions about twice as high received amplification on the order of eight times as strong. It also had fewer explicit follows among the test accounts, making raw Follow counts an inadequate explanation for the difference.

A separate test within one technology publisher found deep-engagement predictions nearly twice as high for accounts that followed the publisher. A United States comparison between ESPN and NFL.com produced a smaller amplification gap of 1.28 times. These were small samples, so none of the figures should become a traffic forecast or universal benchmark. They do support a narrower operational conclusion: learned affinity can materially change distribution even when topic potential is similar, and Follow appears to be one contributing signal rather than a guaranteed reach switch.

You cannot manufacture reader-source affinity with a metadata field. You can, however, make your publication easier for readers and recommendation systems to understand:

• Define a repeatable audience contract. Complete this sentence for each content line: We publish this coverage for this reader at this moment so they can achieve this outcome. If the ending changes radically from one story to the next, the content line may be too diffuse.
• Build continuity, not isolated hits. Connect breaking stories to explainers, updates, recurring series, and logical follow-ups. Item-to-item retrieval and learned source relationships both make continuity more strategically useful than a pile of unrelated traffic bets.
• Protect expectation accuracy. A headline can attract a broad audience that the body was never designed to serve. That may improve the attention layer while weakening evidence of a durable user-source fit.
• Use Follow as reinforcement. Invite readers to follow when you can name the continuing benefit they will receive. Treat the action as an affinity input, not a promise that every follower will see every story.
• Analyze cohorts rather than article averages. Compare returning readers with unfamiliar readers, and compare established coverage areas with occasional topics. A single sitewide average can hide the audience-source combinations that consistently work.

This does not mean your publication must stay inside one narrow subject forever. It means expansion should have a reader bridge. When you enter an adjacent topic, explain why it matters to the audience you already serve and create enough connected coverage to establish a recognizable promise. A one-off article aimed at an unrelated trend may earn attention without building the relationship that supports future distribution.

A practical Google Discover diagnosis FAQ

Why did a strong page receive almost no Discover distribution?

First distinguish low exposure from low response. If the page received few meaningful impressions, you do not yet have a clean headline test; the card had too little opportunity to win attention. Examine whether the story matches a known audience interest, whether its timing fits an enduring or short-term need, and whether it belongs to a recognizable coverage cluster. Because Discover is personalized, absence from one person’s feed is not proof that the page failed retrieval everywhere.

Why did impressions increase while clicks stayed weak?

The page may have entered a candidate pool but failed to earn attention, or it may have been retrieved for people who were not a good fit. Segment the response by topic, reader cohort, and content line before rewriting the title. Then test card packaging that clarifies the subject and payoff without making the promise broader than the page.

Why did clicks rise while reading depth fell?

You likely improved the attention layer without improving the user-content match. Compare the card’s promise with the first screen and the page’s actual depth. Put the central answer or development earlier, remove generic setup, and ensure the rest of the page delivers what caused the click. Continue tracking post-click behavior during packaging tests so a higher click rate does not disguise a weaker experience.

Does asking readers to Follow improve Discover reach?

Follow can contribute to affinity, but it does not guarantee distribution. The strongest time to ask is when a reader has just received value and you can state what future coverage will continue that value. A generic request adds less strategic clarity than an invitation tied to a recurring subject, update cycle, or series.

For your next Discover review, build one funnel view: meaningful exposure, card response, post-click depth, and the difference between returning and unfamiliar readers. Fix the first weak transition instead of blending retrieval, packaging, content quality, and audience strategy into one vague Discover problem.

References

• Search Engine Land — Google’s Discover job posting confirms what the feed already shows