Google Discover looks like a black box from the outside, but it behaves predictably enough that a structured analysis can reliably surface more value from it. The platform rewards fresh, entity-driven content that earns early engagement, and once you understand that mechanic, you can build a repeatable process around it instead of chasing whatever happened to go viral last week.

None of this means gaming the feed with curiosity-gap headlines about miracle cures. High click-through rate wins the first few hours; user satisfaction decides whether a publisher keeps its Discover presence at all. The goal here is a framework for measuring what actually drives performance, not a shortcut around it.

What Discover actually rewards

A few mechanics define how the feed behaves, and they’re worth internalizing before you open a spreadsheet:

• Discover runs on freshness. Evergreen content does appear, but it tends to ride alongside something newsworthy rather than stand alone.
• Lifestyle and engaging formats generally outperform dry utility content.
• Like news surfaces, Discover is heavily entity-, click-, and early-engagement-driven.
• The feed groups readers into personalization cohorts. Satisfy one member of a cohort and more of that cohort tends to follow.
• Content that beats its own predicted early performance is more likely to get amplified further.
• Once the interested cohort is saturated, performance drops off naturally rather than gradually.

Most recommendation algorithms, Discover included, follow something close to a golden-hour rule: the first sixty minutes after publishing largely determine whether the system amplifies a piece or lets it sink. If early sharing and engagement are strong, push them immediately — that window is where the algorithm makes its first real judgment call.

The seven data points worth tracking

This analysis assumes you’re already pulling basic conversion, click, and impression data from Discover — that’s the baseline, not the destination. Treat these seven areas as the layer above it. Also worth knowing upfront: Discover traffic isn’t cleanly separable in most analytics platforms. Teams typically approximate it using a combination of Google referral and mobile/Android signals, and accept that some of it will always be misattributed.

Click-through rate

CTR is foundational to Discover, Top Stories, and any real-time SEO surface, more so than in traditional organic search, because the algorithm is deciding what to promote in something close to real time. It’s weighted alongside broader engagement signals — clicks, on-page behavior, session duration — that connect a clickable headline and image to content that actually satisfies the reader. That combination is also why pure clickbait has been losing ground: even readers who click on hype eventually stop trusting the source. To get real signal from CTR, pair it with image type, headline type, and entity data rather than reading it in isolation.

Entity analysis

Entities matter more in news and Discover than almost anywhere else in search. People, places, and organizations account for the large majority of entities worth tracking, and understanding which ones drive engagement for your content is one of the highest-value exercises you can run.

Doing this well requires more than eyeballing headlines. A practical setup combines an LLM, a named entity recognition tool, and either Google’s Knowledge Graph or Wikidata: extract candidate entities from the title, disambiguate them against the page content (so “apple” resolves correctly to the company, the fruit, or a celebrity’s child), then confirm the match against a knowledge base. Once you have clean entity data, a simple bubble chart plotting volume against performance is usually enough to spot content opportunities at a glance.

Subfolder performance

On larger sites with heavy publishing volume, tracking impressions and clicks by subfolder shows you where Discover concentrates its attention. This becomes far more useful once you layer in headline type and entity data on top, because it lets you tell commissioners and writers which formats tend to work in which sections rather than giving generic advice. Pull total clicks, impressions, and CTR per subfolder, along with an average per article, and track the trend over time.

Authorship signals

Google tracks authorship as a search signal, and it factors into E-E-A-T in ways that aren’t fully disclosed but are clearly meaningful — leaked internal documentation on ranking signals showed a large share of author-related categories dealt specifically with identifying who wrote a piece and how clear their footprint was online. Disambiguation matters across the board right now, from structured data to knowledge graphs, largely as a defense against synthetic content and misinformation.

For Discover specifically, evaluate authors on three questions: how many of their articles make it into Discover and perform in Search, which topics and entities they perform best with, and whether certain headline types work better for them than others.

Headline type

This is where you learn what kind of content actually earns clicks from your audience. Do curiosity-gap headlines outperform straightforward ones? Do numbered headlines beat prose headlines? Does a celebrity name in the headline move CTR, and does that effect vary by section? Do first-person headlines perform differently in a finance section than in news?

You can’t scrape Discover directly, but you can build a defensible hypothesis by categorizing every headline you publish — curiosity gap, localized, numbered list, question, how-to, emotional trigger, first person — and measuring CTR against each category using a language model trained on your archive. Verify the categorization manually before trusting it, then break the results down by subfolder, author, or entity. It’s also worth remembering Discover tends to lean on the Open Graph title more than traditional search does, which gives you room to run a punchier headline there than you would in the H1 or page title.

Image testing

Images function much like headlines: impossible to prove definitively which one Discover pulls, but a featured image at 1200px wide is a safe assumption for the one in use. CTR is the dominant factor in early-stage success; ongoing performance leans more on traditional engagement signals. Two controllable levers decide Discover CTR — the headline and the image.

Testing consistently shows that in hard news, images of people showing visible emotion, especially sadness or distress, tend to outperform neutral shots. In finance content, the opposite holds: people looking directly at the camera and appearing confident or happy perform better, likely because it reduces the perceived risk of a financial decision. Publishers also test badges and logos on thumbnails — they reliably lift CTR for recognizable brands, and if a paywalled site runs a free live blog, flagging that in the image is worth the effort.

Break image analysis into human presence and gaze, facial expression, emotional tone, composition, color palette, and photo type, then use a model to bucket images into groups — camera-facing and smiling versus off-camera and neutral, for example — to see which buckets actually move CTR.

Publishing cadence

This matters in proportion to publishing volume. If you’re placing 50 or fewer articles into Discover per month, this layer probably isn’t worth the effort yet. At hundreds or thousands of articles, patterns in publishing day, publishing time, content freshness, and republishing versus fresh publishing become genuinely useful guidance for commissioning desks.

Turning the analysis into action

Set clear goals before you start pulling data — specifically around what you can actually influence. You may not control what a desk chooses to publish, but you can usually change the headline and image before it goes live, which is where most of the controllable CTR lives.

It also helps to be honest about whether your role is strategic or tactical. A strategic role advises on the entities and headline types worth pursuing without necessarily implementing them directly. A tactical role has direct control over headlines, publish timing, and entity targeting. Knowing which one you’re in determines whether your output should be a recommendation memo or a direct edit.

One caution worth keeping in view: avoid making decisions purely for Discover’s benefit. It’s a volatile platform that can reward thin, curiosity-driven content if you optimize for its numbers in isolation, and it has limited direct impact on revenue compared with core organic search. Use it as one input among several, not the target itself. For the broader mechanics behind how a feed like this actually surfaces content, see our breakdown of how recommender systems like Google Discover may work, and for the wider shift in who benefits from the feed, why Google Discover is no longer just for publishers. The entity groundwork underneath all of this is covered in how to implement entity optimization without relying on schema markup, and thumbnail selection specifically is addressed in how Google picks thumbnails for Search and Discover.

Frequently asked questions

What is the most important metric for Google Discover?

Click-through rate drives early-stage amplification more than anything else, but it needs to be read alongside entity performance, headline type, and image type rather than tracked alone — and sustained success depends on satisfying readers, not just earning the click.

Can you accurately track Discover traffic in analytics tools?

Not cleanly. Most platforms don’t isolate Discover as its own channel, so teams approximate it using Google referral traffic combined with mobile or Android signals, accepting that some of it will be misattributed.

Why do entities matter so much for Discover performance?

Discover behaves similarly to news surfaces, which are unusually entity-driven. Understanding which people, places, and organizations in your content correlate with strong performance lets you make more informed decisions about coverage and headline framing than keyword-level analysis alone allows.

Should content be created specifically to perform in Discover?

Generally no. Discover is volatile and can reward thin, curiosity-driven content in the short term, but it has limited direct revenue impact. Use Discover analysis to refine headlines and images on content you’d publish anyway, not to dictate what gets commissioned.