How Featured Snippets Are Chosen for the Answer Box

A search engine picks a featured snippet by extracting one passage from a page that already ranks in the top results for that query, taking the passage that answers the query most directly, in the format the query shape calls for. There is no markup that opts a page in and no way to nominate a passage. Selection is recomputed every time the query is run.

Everything below is the detail behind that paragraph: which formats exist and what pulls each one, why the passage matters more than the page, what "already ranks" implies for where your effort should go, and why the box is not always worth winning.

A featured snippet is a block of text lifted from a single web page and displayed above the standard results, with the source page's title and link attached underneath it. People call it position zero. That nickname is informal and no search engine uses it, which matters more than it sounds, because "position zero" implies an extra slot that gets added to the page. It is not an extra slot.

Google's documentation describes the mechanism in one sentence: "Google systems determine whether a page would make a good featured snippet for a user's search request, and if so, elevates it." Read the verb. The listing is elevated, not duplicated. Since January 2020 a page elevated into the box does not also appear again in the ordinary page one results. You swap your blue link for the box. The full comparison of that behavior against AI answer panels, including how the two surfaces get counted differently in reporting, lives in AI Overviews vs featured snippets, and this post stays on the selection side.

One line from the same documentation is worth more than it usually gets: "Clicking a featured snippet takes the user directly to the section of the page that appeared in the featured snippet." The visitor does not land at the top of your article. They land at the extracted passage, mid page, with your introduction above them and no context. Whatever section wins the box is the section that has to work as an arrival point.

Other search engines run comparable answer boxes with broadly similar rules. The specifics quoted throughout this post come from Google's published documentation, because that is the set that is actually written down.

The box takes four shapes, and the query decides the shape, not you. You cannot answer a comparison query with a paragraph and expect a table.

Ordered versus unordered matters inside the list format. Sequence queries ("steps to", "how to", "process for") pull ordered lists. Set queries ("types of", "reasons for", "what to include") pull unordered ones. Getting that backwards is a common reason a well-written section never gets picked: the content is right and the container is wrong.

This is the mechanic that changes how you plan the work. Your page does not have to be about the query. It only has to contain a passage that answers it.

A 3,000 word guide to email deliverability can win the box on "what is a soft bounce" using 55 words buried in section seven. The page's title, its primary keyword and its overall topic are inputs to whether the page ranks at all. They are not what gets extracted. What gets extracted is a block.

Two separate systems are in play here, and conflating them produces bad advice:

• Extraction has always been passage level. The box shows a section of a page. That has been true since the feature launched.
• Ranking now has a passage level component too. Google's ranking systems guide describes it as "an AI system we use to identify individual sections or 'passages' of a web page to better understand how relevant a page is to a search."

The useful consequence is that snippet opportunities are frequently sitting inside content you published for something else entirely. Before planning a new page, audit the long ones you already have for buried answers. That is a one-paragraph rewrite, not a content project.

The uncomfortable consequence is the mirror image. Your page can be represented in the box by a passage you would never have chosen: a definition you wrote loosely in an aside, a caveat, a line from a callout. You have no way to nominate a different one. You can only remove candidates, which the section on markup covers.

Ahrefs published a study in May 2017 covering 2 million featured snippets drawn from roughly 112 million US keywords. Its headline finding was that 99.58% of featured snippets came from pages already ranking in the top 10. The same study found that only about 30.9% of snippet sources were the number one result, and that around 12.29% of queries in the dataset showed a snippet at all.

Two caveats before anyone builds a strategy on those decimals. The study is from 2017, so it predates the January 2020 deduplication change and predates AI answer surfaces entirely. And it is vendor research, not official documentation. What has held up is the direction, and the mechanism explains why it should: the box promotes a result that already exists. If your page is not in the result set for that query, there is nothing to promote.

Treat the top-10 finding as a filter, not a tactic. Formatting is the tiebreaker among pages that already qualify. Almost every guide on this subject sells formatting as a way to leapfrog from anywhere, which quietly inverts the effort order and sends people off to rewrite paragraphs on pages that were never candidates.

Here is the order that actually reflects the mechanism:

That last row is the only one where the decision is genuinely close, and the section on whether winning is a win covers it.

Extractability is a property of the block, not of the page. Five things decide it.

A heading shaped like the question. Nothing requires your heading to match the query, and plenty of snippets get pulled from under headings that do not. What a question-shaped heading gives the extractor is an unambiguous boundary: here is where the answer starts, here is where it stops. Ambiguous boundaries mean the engine has to guess, and guesses favor pages that did not make it guess.

The answer in the opening sentence. The box truncates from the end, so a windup gets shown and the answer gets cut. Write the definition or the conclusion first, then the context beneath it. That is the same discipline that makes a page work above the fold, covered in satisfying search intent above the fold.

A length that fits the box. No search engine publishes a limit. Open the box you are trying to win, count its words, and write to roughly that. Around 50 words is a reasonable default for a paragraph target.

Markup that matches the format. Real list elements for list queries. A real table with a header row for comparison queries. Sequential subheadings work well for step lists, because the engine can read them as items.

Self-containment. The passage has to name its own subject. A block opening with "It also affects deliverability" is meaningless once it is cut out of the page, and a passage that is meaningless alone does not get shown alone.

None of this is a lever you pull on the search engine. It is the removal of reasons not to pick you.

People arrive at this topic looking for the markup. There isn't any. No FeaturedSnippet type exists in schema.org, there is no meta tag, no Search Console submission and no application. Google's own help documentation answers the question "how can I mark my page as a featured snippet" with two words: "You can't."

Structured data does real work elsewhere: eligibility for rich results, entity clarity, machine readability for answer engines. That work is worth doing, and schema markup for beginners covers what it actually earns. It is not a snippet application, and any guide implying FAQPage or Article markup improves snippet odds is guessing.

The only markup that reliably changes snippet behavior is the set that takes you out:

Look at the shape of that table. You have a reliable veto and no reliable nomination. That is the entire control surface, and it is worth internalizing before spending a month on snippet optimization.

data-nosnippet has one genuinely positive use that rarely gets mentioned. If the engine keeps extracting the wrong block from your page, a promotional line, a disclaimer, an author bio, a cookie notice that leaked into the content area, wrapping that block removes it from the candidate pool. You are not choosing the winner. You are removing a loser, and sometimes that is enough.

The cheapest snippet work is on queries where you already rank in the top 10 and the answering passage is badly shaped. Here is how to find those, with the exact path and settings. Every number below is illustrative, not SparkCliks data.

Step 1. Open Google Search Console, then Performance, then Search results.

Step 2. Set the date range to Last 3 months. Twenty-eight days is fine for measuring a change but too thin for discovery, because the long-tail question queries you want are the ones with the fewest impressions.

Step 3. Add a Query filter set to Custom (regex) and use a pattern that catches question shapes:

^(what|why|how|when|who|which|is|are|can|does|do|should)\b

Search Console regex uses RE2 syntax, so this works as written. It over-matches slightly, which is fine at the discovery stage.

Step 4. Enable all four metrics (Clicks, Impressions, CTR, Average position) and export the Queries tab.

Search Console lets you sort by average position but does not let you filter on it, so position filtering happens after export. This is the step most walkthroughs skip, and it is why people report that the report "does not have that filter". It does not. The spreadsheet does.

Step 5. In the export, keep rows where average position sits between 2 and 10 and impressions clear a floor that fits your site. As an example, 100 impressions over three months on a small site, or 1,000 on a large one. Sort by impressions descending.

Step 6. Take the top 20 rows and run each query yourself in a private window. Record three things per query: whether a box appears at all, which of the four formats it is, and who owns it. Expect most of your list to have no box. Those rows are dead ends for this project, and checking first is what stops you rewriting 20 paragraphs for nothing.

Step 7. For the rows that do have a box, compare your answering passage against the one being shown. In practice the gap is almost always one of the six rows in the extractability table above: the answer is buried, the block opens with a pronoun, the paragraph runs triple the box length, or the format does not match.

Two things work against the box on informational queries, and both are structural rather than bad luck.

The first is that a complete answer removes the reason to click. If the searcher wanted a definition and the box gave them the definition, the session is over. That is the zero-click pattern, and a snippet is one of its most efficient engines. The better your answer, the more completely it ends the visit.

The second is deduplication. You are not adding a box on top of your listing. You are trading the listing for the box.

About the numbers people quote here. Ahrefs' 2017 study reported that on result pages carrying a snippet, the box took roughly 8.6% of clicks while the first ordinary result beneath it took roughly 19.6%, against roughly 26% for a number one ranking on a page with no snippet. Those figures get repeated constantly, almost always without the caveat that they were collected in May 2017, nearly three years before deduplication arrived. In that dataset a snippet source could still appear in the list as well, so the click split it describes is not the split you face now. No comparable post-2020 measurement has been published. Do not use 8.6% as a forecast, and be skeptical of anyone who does.

What survives the caveat is the direction, and it is enough to decide with:

• The box does not automatically out-click a strong ordinary listing.
• You can no longer hold both, so every snippet win on a query where you already rank well is a trade rather than an addition.
• On a query where a competitor holds the box and you sit at 2 or 3, taking it is usually right, because the alternative is that they keep it.

Ask one question per query before spending the effort: after reading the box, what does the searcher still need? If the answer is "nothing", winning it may cost you traffic. If the answer is "the other seven rows of that table", or "the version for their situation", or "the tool", the box works as an advertisement for the click. Comparison and specification queries tend to sit in the second group. Bare definitional queries tend to sit in the first.

Check organic CTR benchmarks for what your click share should already look like at your current position, so you know what you are risking. And for the wider picture of how answer surfaces are eating the click, including how AI panels behave differently from snippets on exactly this point, read the AI Overviews and featured snippets comparison.

Snippet ownership is not a state you hold. It is an outcome recomputed for every query, every time, and it is volatile in a way rankings usually are not. Pages gain and lose boxes on the same query inside a week with nobody editing anything.

When a box disappears, four different things might have happened, and the right response differs for each:

1. A competitor's passage now reads as the better answer. They rewrote a paragraph, or a new page entered the top 10. This is the only case where rewriting yours is the correct response.
2. The box stopped firing on the query entirely. Nobody owns it now. The result page changed shape, which happens constantly as query interpretation drifts. Nothing was lost to a competitor and nothing needs rewriting.
3. The interpretation of the query shifted. The engine now reads it as a different intent, which changes the format it wants. Your paragraph is fine and the box now wants a table.
4. A policy removal. Google publishes content policies specific to featured snippets covering dangerous content, deception, harassment, hate, medical misinformation and similar categories, and it removes boxes that violate them. Those policies apply to the box, not to your ordinary listing.

Run the query in a private window before doing anything. Cases 2 and 3 account for a large share of "we lost our snippet" panics, and both get misdiagnosed as case 1 constantly.

Two further notes. Rank tracking tools disagree with each other about snippet ownership because they sample different locations, devices and times, so a tool alert is a prompt to check rather than a fact to act on. And the right way to budget this work is as a portfolio: format 30 answering passages well and expect a shifting subset of them to hold boxes at any moment, rather than defending one query as though it were an asset. It is not an asset. It is a recurring decision you do not get to participate in.

Run this against a specific query before touching a page.

1. Does the query show a box today? Check in a private window first. If there is no box, stop here.
2. Is the page already in the top 10 for that query? If not, formatting is premature and the ranking work comes first.
3. Which of the four formats is the box using? Match it. A paragraph will not win a table query.
4. Does your answering passage give the answer in its first sentence, before any context?
5. Does the block name its own subject, rather than opening with "it" or "this"?
6. Is it roughly the length of the box you are looking at, rather than three times it?
7. For step and set queries, is there real ordered or unordered list markup, or parallel subheadings?
8. For comparison queries, is there a real table with a header row?
9. Is a boilerplate block getting pulled instead of your answer? If so, consider data-nosnippet on that block.
10. If you win it, does the reader still have a reason to click through? If not, decide deliberately whether you want it.
11. Did you record a baseline and pick untouched control pages before making the edit?

Nothing in that list is a guarantee. Selection is the search engine's call, it gets recomputed every time, and no amount of formatting entitles you to the box. What the list does is remove the reasons a qualified page gets passed over, which is the only part of this you control.