
How to Get Featured Snippets and AI Citations
On this page
SEO
Most advice treats featured snippets, People Also Ask and AI citations as separate games requiring separate tricks. If you want to know how to get featured snippets, start with one useful passage on an indexed, snippet-eligible page. Google requires an AI Overview supporting page to be “indexed and eligible to be shown in Google Search with a snippet.”
Google’s AI features documentation specifies no additional eligibility requirements. Its featured snippets documentation says snippets can appear within a related questions group (People Also Ask). You cannot mark a page as a featured snippet; Google chooses it.
The controls overlap too: nosnippet prevents snippets and direct input for AI Overviews and AI Mode; max-snippet limits both. data-nosnippet excludes marked text from snippets, including what AI features show. One liftable passage is therefore a sensible publishing unit across these surfaces, although eligibility never guarantees selection.

The same passage can supply several answer surfaces. Snippet restrictions affect those opportunities together, although excluding a marked block with data-nosnippet has a narrower effect than blocking the whole page with nosnippet.
Choose controls before editing answers. A well-written paragraph cannot overcome an instruction telling Google not to show it.
| Goal | Control | Effect |
|---|---|---|
| Discourage featured snippets while keeping regular snippets | Lower max-snippet stepwise | May work; Google provides no guaranteed character threshold |
| Block Google snippets and direct AI Overview or AI Mode input from a page | nosnippet or max-snippet:0 | Blocks snippets and that direct input; does not remove the page from the index |
| Exclude one block from snippets | data-nosnippet on span, div or section | Excludes marked text from snippets, including what AI features show |
| Exclude the site from Google Search generative AI features | Search generative AI control in Search Console | Site-wide control; check your property’s Settings |
| Remove a page from Google Search | noindex | Removes the page from results after Google processes the rule |
| Exclude the site from ChatGPT search answers | Disallow OAI-SearchBot | Approximately 24 hours to adjust; navigational links can remain |
| Exclude the site from Perplexity search retrieval | Disallow PerplexityBot | Changes can take up to 24 hours |
| Restrict Bing conversational answers to URL, title and snippet | NOCACHE | Bing’s stated September 2023 policy permits that limited use |
| Exclude content from Bing conversational answers and training | NOARCHIVE | Bing’s stated September 2023 policy keeps tagged pages eligible for regular search |
Google documents no per-page control that keeps featured snippets while excluding the page from AI Overview input. OpenAI’s GPTBot is a training crawler, separate from OAI-SearchBot; user-initiated ChatGPT-User requests may not follow robots.txt rules. Perplexity describes PerplexityBot as a search crawler, not a foundation-model training crawler.
The MUST requirements: eligibility and extractable answers
M1. The page permits the use you want
Keep the page indexed, crawlable and eligible for snippets when Google answer visibility is your goal. Google must fetch the page to read its robots directives; blocking it in robots.txt prevents those page-level instructions from being read.
The following are alternative controls, not a recommended combined configuration. Sending both meta rules would leave the more restrictive instruction in charge.
<!-- page level: no snippet anywhere in Google Search, incl. AI Overviews and AI Mode input -->
<meta name="robots" content="nosnippet">
<!-- page level: cap the text snippet at 50 characters (also limits AI Overviews / AI Mode input) -->
<meta name="robots" content="max-snippet:50">
<!-- text level: keep one block out of snippets; only span, div, section; value is ignored -->
<div data-nosnippet>Internal pricing note that should not be quoted.</div>
The HTTP header equivalent is X-Robots-Tag: nosnippet. Google’s robots directive documentation specifies that max-snippet counts characters: 0 equals nosnippet, while -1 lets Google choose.
Use data-nosnippet only on valid, closed span, div or section elements. Any attribute value is ignored, so data-nosnippet="false" still excludes the text. Never add or remove the attribute on existing nodes through JavaScript: extraction can happen before or after rendering.
Google’s featured snippets documentation says nosnippet takes priority over data-nosnippet. The robots documentation separately establishes that conflicting directives resolve to the most restrictive rule.
URL="https://example.com/your-page"
curl -sI "$URL" | grep -i '^x-robots-tag'
curl -s "$URL" | grep -oiE '<meta[^>]+name="(robots|googlebot)"[^>]*>'
curl -s "$URL" | grep -c 'data-nosnippet'
Empty output from the first two checks suggests default snippet permission; confirm that both requests succeeded. This quick check is not a complete HTML parser and can miss differently formatted tags. The last command counts lines mentioning data-nosnippet, not attributes or elements.
Why: Google documentation establishes the eligibility gate and directive behavior. Formatting cannot override a restrictive rule.
Test: Pass when Search Console confirms indexing eligibility and fetched headers, HTML and robots.txt permit the intended use. Fail when nosnippet, max-snippet:0, noindex or a crawl block contradicts that goal.
M2. The answer comes first under its heading
Write the direct answer in the first paragraph under the relevant heading. Give its scope and essential qualification there, then expand with evidence or instructions. A reader should not need the company history before discovering the replacement cost.
The familiar length recommendation comes from old descriptive research. A.J. Ghergich’s January 2017 Moz/Semrush analysis examined 1.4 million snippets across 30 question, preposition and comparison phrases, reporting paragraph answers around 40 to 50 words or 300 characters.
A November 2020 Search Engine Journal summary of Semrush research reported paragraphs at 70% of snippets, averaging 42 words and 249 characters. Both studies predate AI Overviews; neither establishes a causal word-count threshold.
Pew’s July 2025 research found a 67-word median among 12,593 AI summaries collected by rerunning participants’ March searches in April 2025. That measures complete generated summaries, not the ideal length of a source paragraph.
Growth Memo reports that 44.2% of 18,012 verified ChatGPT citations came from the first 30% of a page. Its free library description states the finding, but the full study is paywalled. Treat it as observational evidence about ChatGPT, not proof about Google or a controlled placement experiment.

Growth Memo’s reported ChatGPT citation distribution favors the beginning of a page. It supports putting the main answer early, without establishing a placement rule for every assistant.
Why: Historical snippet studies and Growth Memo’s citation analysis are descriptive evidence. They support accessible opening answers, not guaranteed extraction lengths.
Test: Pass when the first paragraph answers its heading without requiring earlier context. Fail when it opens with background, promotion or an unanswered question.
M3. The answer exists in server HTML and is immediately visible
Put the answer in the HTML returned by the server and display it without interaction. A fetcher that does not execute JavaScript cannot extract text that appears only after browser code runs.
Check both conditions separately. Text can exist in source while remaining hidden in a closed accordion; browser-visible text can also be absent from the initial response. The server-response checks in our crawl-cost guide help inspect what a non-rendering fetcher receives.
Google’s snippet documentation added read-more deep-link guidance on 20 April 2026. It recommends immediately visible content, avoiding expandable sections or tabs, keeping the hash fragment, and avoiding JavaScript that controls the user's scroll position on page load.
Featured snippet clicks can also scroll automatically to the extracted passage without special annotation. Unsupported browsers or an undetermined passage location can produce a landing at the page top. Make both landings usable.
Why: Google documentation supports immediately visible deep-link destinations; server-response inspection tests accessibility to non-rendering fetchers.
Test: Pass when an exact answer sentence appears in fetched HTML and is visible on initial load. Fail when it needs JavaScript insertion, a tab click or expansion; we treat closed <details> as failing this house check.
M4. Every answer survives being lifted out of context
Name the subject and include the condition that changes the answer. “It takes three days” becomes ambiguous when detached from the heading or preceding paragraph. The passage should identify what takes three days and under which assumptions.
A hypothetical water-heater page could place its dated price range and installation assumptions directly under the cost heading, followed by line items. No actual price or outcome is implied here. The structural requirement is that the range travels with its qualifications.
On public-interest topics, Google’s featured snippet policies say the snippets it shows should not contradict well-established or expert consensus. This governs Google’s snippet selection; it is not a requirement imposed on pages. If a website has many snippets that break the rules, Google stops showing featured snippets from that site.
Why: Google documentation establishes consensus rules for displayed featured snippets and site-wide policy enforcement. Self-contained wording is our editorial requirement for preserving meaning during extraction.
Test: Pass when a reviewer can read the answer alone and identify its subject, scope and qualifications. Fail when extraction changes its meaning or leaves unsupported consensus-sensitive advice.
M5. The answer-first check runs against the published response
Run this standard-library Python script against the public URL after publishing. It checks the first paragraph following each H2 or H3 in raw server HTML, without running JavaScript.
import re, sys, html, urllib.request
url = sys.argv[1]
req = urllib.request.Request(url, headers={"User-Agent": "Mozilla/5.0 answer-first-check"})
raw = urllib.request.urlopen(req, timeout=20).read().decode("utf-8", "ignore")
raw = re.sub(r"(?is)<(script|style|noscript)\b.*?</\1>", " ", raw)
def text(s):
return re.sub(r"\s+", " ", html.unescape(re.sub(r"<[^>]+>", " ", s))).strip()
# each h2/h3, then the first <p> that follows it in the raw server HTML (no JavaScript run)
for m in re.finditer(r"(?is)<h([23])\b[^>]*>(.*?)</h\1>(.*?)(?=<h[23]\b|$)", raw):
level, head, body = m.group(1), text(m.group(2)), m.group(3)
p = re.search(r"(?is)<p\b[^>]*>(.*?)</p>", body)
first = text(p.group(1)) if p else ""
words = len(first.split())
flag = "OK" if 0 < words <= 60 else "CHECK"
print(f"h{level} | {words:3d} words | {flag} | {head[:60]}")
Save it as answer_first.py, then run python3 answer_first.py followed by your URL. The recorded output from Google’s featured snippets page on 24 October 2026 was:
h2 | 12 words | OK | How can I opt out of featured snippets?
h3 | 23 words | OK | Block all snippets
h3 | 45 words | OK | Block featured snippets only
h2 | 24 words | OK | How can I mark my page as a featured snippet?
h2 | 79 words | CHECK | What happens when a user clicks a featured snippet?
h3 | 0 words | CHECK | Get support
h3 | 0 words | CHECK | Resources
h3 | 0 words | CHECK | Tools
The 60-word flag is our house threshold, not a Google number. The 79-word answer can still be useful, while the zero-word footer headings are irrelevant noise. An answer available only after JavaScript runs can produce zero words or the wrong paragraph.
Why: This is a mechanical house check informed by descriptive length studies. It detects missing or delayed answers, not ranking potential or factual quality.
Test: Pass when every relevant heading has a direct answer and every CHECK has been reviewed. Fail when missing server-rendered answers remain unexplained.
For a company whose answers exist but sit buried, hidden in tabs or rendered only in the browser, our AI visibility service audits buyer questions and rebuilds pages around accessible answers. The work targets observable publishing defects; it promises no snippet or citation count.
The SHOULD requirements: useful formats and defensible evidence
S6. The format matches the information
Use a paragraph for a definition or qualified answer, a numbered list for a sequence, and a table for comparable fields. Select the structure a reader needs before considering how a search feature might display it.
The 2020 Semrush study summarized by Search Engine Journal reported 19.1% lists, averaging six items and 44 words, and 6.3% tables, averaging five rows and two columns. Videos accounted for 4.6%; paragraphs accounted for the remaining 70%. These are historical format shares, not recommended quotas.
A paragraph-length target does not govern tables or lists. A four-row comparison should keep its meaningful labels and units even when those exceed a paragraph’s character count.
Why: Historical descriptive research shows multiple snippet formats. Semantic HTML gives each information type an appropriate structure.
Test: Pass when steps have an order and table headers identify comparable values. Fail when prose disguises a comparison or a list exists solely to chase a snippet.
S7. Evidence improves the answer without becoming decoration
Add statistics, quotations and cited sources where they resolve uncertainty. Preserve the sample, date and limitations next to striking numbers so an extracted sentence does not turn a narrow result into a universal claim.
The GEO paper’s “up to 40%” finding is a lab result, not a forecast for your website. Its KDD 2024 work used a 10,000-query benchmark, a 2023-era generative engine and a Perplexity test.
Against a baseline position-adjusted word count of 19.5, quotation addition scored 27.8, statistics 25.9 and cited sources 24.9: approximately 43%, 33% and 28% higher. Cite Sources gained 115.1% at source rank five but lost 30.3% at rank one, showing that starting position changed the result.
This separate, broader benchmark tested multiple tasks, domains and numbers of competing adopters. Many conversational optimization methods were largely ineffective or harmed document ranking; gains shrank as more competitors adopted them, while improving source placement in the model context was more effective.
These benchmarks ask different questions and should not be combined into one promised uplift. A quotation earns its space by supplying evidence, not by being quotation-shaped.
Why: Lab benchmarks offer bounded evidence about particular systems and competitive settings. They do not establish live-search causal effects for a company page.
Test: Pass when each added statistic changes a decision and retains its source context. Fail when decorative numbers or quotations support a promised citation increase.
S8. Long pages keep clear sections and substantive updates
Keep a complete page when the subject requires it, with headings that identify each answer. Passage ranking gives Google a way to evaluate relevant sections inside longer pages; it does not require turning every paragraph into a separate URL.
Google announced passage understanding on 15 October 2020, projecting improvements for 7% of queries across languages during global rollout. Trade coverage reported a US English launch on 10 February 2021 and explained the rename from passage indexing to passage ranking: passages were not being separately indexed.
Freshness also depends on the question. Ahrefs’ 28 July 2025 analysis covered approximately 17 million citation and result URLs: AI assistants’ cited URLs averaged 1,064 days old against 1,432 for organic results. AI Overviews were measured separately and sat close to organic results.
SE Ranking’s 24 November 2025 ChatGPT study covered 216,524 pages across 129,000 domains and 20 niches. Recently updated pages averaged 6.0 citations against 3.6 for outdated pages, but correlated variables and its observational model prevent a causal conclusion.
Why: Google documentation supports passage relevance and query-dependent freshness. Citation-age studies provide correlations, not a universal update schedule.
Test: Pass when headings expose distinct answers and revisions record substantive changes. Fail when a long page hides its answer or changes dates without changing facts.
S9. Organic relevance remains part of the citation plan
Continue improving relevance and source quality alongside extractability. A page need not rank in the organic top ten for the visible question to appear in an AI Overview; Ahrefs found most citations came from outside those positions.
Ahrefs’ March 2026 study examined 863,000 keyword SERPs and four million AI Overview URLs. Counting organic links only, 37.1% of citations ranked in the top ten, 26.2% in positions 11 to 100 and 36.7% outside the top 100. These categories describe the visible query, not every related retrieval query. Ahrefs suggests query fan-out as a likely explanation, which Google says AI Overviews and AI Mode may use, and notes that some cited pages appear only in SERP features.
Why: Google documentation describes possible fan-out; Ahrefs supplies observational rank distributions and a proposed explanation. Neither establishes fan-out as the cause or supports abandoning organic relevance.
Test: Pass when the page answers its target question and the related subquestions it introduces. Fail when a snippet-shaped opening substitutes for useful supporting content.
The AVOID list: unsupported shortcuts and misleading signals
A10. Avoid treating FAQ markup as an answer-box switch
FAQ rich results stopped appearing from 7 May 2026, and Google removed their documentation on 15 June 2026. Search Console’s performance help distinguishes explicitly coded FAQ rich results from algorithmic People Also Ask suggestions; featured snippets remain selected by Google.
General schema markup has a different job. Ahrefs’ 1,885-page controlled comparison matched pages adding JSON-LD against 4,000 controls, using 30-day before-and-after windows between August 2025 and March 2026.
The difference-in-differences results were AI Overviews down 4.6%, AI Mode up 2.4% and ChatGPT up 2.2%. Only the small AI Overview decrease was statistically significant; the study found no clear general citation benefit, with limits including already heavily cited pages and JSON-LD-only measurement.
Use our structured-data guide to choose markup for supported rich results. Google says structured data is not required for generative AI search.
Why: Google documentation describes feature eligibility; Ahrefs’ controlled observational test weakens the claim that adding schema causes more citations.
Test: Pass when markup supports a documented feature and matches visible content. Fail when FAQ or generic JSON-LD is sold as a citation trigger.
A11. Avoid making llms.txt a visibility project
Google’s AI optimization guidance says Search does not need special AI text files or machine-readable formats. Treat llms.txt as hygiene at most, never a visibility lever.
Ahrefs examined 137,210 domains using its analytics product in May 2026. Of approximately 38,000 valid files, 97% received zero requests that month; AI retrieval bots accounted for about 1.1% of requests to these files. The sample skews technical, and fetching a file would not prove using it.
Why: Google documentation rejects the requirement; Ahrefs’ request-log study provides usage evidence with sampling limitations.
Test: Pass when answer accessibility work proceeds independently of this file. Fail when publishing it becomes the proposed remedy for missing citations.
A12. Avoid arbitrary chunks and compulsory question headings
Write sections around complete ideas and use questions when questions describe their contents. Google explicitly says tiny content chunks are unnecessary and specifies no ideal page length.
SE Ranking’s ChatGPT correlations resist simplistic formatting rules. Sections of 120 to 180 words averaged 4.6 citations against 2.7 for sections under 50 words; sections over 180 averaged 5.7. Question-style headings averaged 3.4 citations against 4.3 for straightforward headings.
These interdependent observations cannot prescribe section lengths or prohibit questions. They do show why “make everything shorter and turn every heading into a question” is poor evidence-based advice.
Why: Google documentation rejects mandatory chunking; SE Ranking’s vendor correlation study supplies no causal heading formula.
Test: Pass when each heading accurately labels a complete answer. Fail when arbitrary limits split explanations or force unnatural questions.
A13. Avoid fake freshness
Update dates when the underlying information changes. Rechecking a price, revising a specification or replacing obsolete instructions gives a reader something useful; a scheduled timestamp rotation does not.
Ahrefs’ freshness findings are correlations, and its report notes Google’s warning against changing publication dates without real changes. Average cited ages measured in years also contradict the idea that only newly published pages qualify.
Why: Citation-age correlation studies support investigating freshness, while Google’s query-dependent freshness systems explain why its value varies.
Test: Pass when a revision log identifies changed facts or instructions. Fail when the only changed field is the date.
A14. Avoid stuffing keywords or manufacturing mentions
Use terminology that identifies the subject accurately and pursue references earned by useful work. Repetition and inauthentic mentions offer no defensible substitute for evidence.
In the GEO lab benchmark, keyword stuffing scored 17.8 against the 19.5 baseline, approximately 8.7% lower. Google clarified on 15 May 2026 that its spam policies also apply to generative AI responses, and its optimization guidance discourages inauthentic mentions.
Why: A lab benchmark measured a negative stuffing result; Google documentation applies spam policies to generative responses.
Test: Pass when repeated terms clarify meaning and external mentions have a legitimate editorial purpose. Fail when repetition or manufactured references exist solely to influence retrieval.
A15. Avoid treating a featured snippet as extra organic inventory
A featured snippet does not preserve a second copy of the listing later on the first results page. Danny Sullivan announced deduplication on 22 January 2020, as reported by Search Engine Journal; Google documents the behavior in its ranking systems guide.
Search Console typically counts a featured snippet at position 1. Compare clicks and qualified visits before and after a verified feature change; an apparent position improvement alone does not establish a business gain.
The opportunity has also changed. Ahrefs’ one-million-SERP comparison found featured snippets falling from 15.41% in January 2025 to 5.53% in June, while AI Overviews rose from 3.93% to 27.43%.
The question feature, which we read as People Also Ask style boxes, rose from 73.73% to 79.13%. Featured snippet presence correlated at -0.9053 with AI Overview growth; this does not prove replacement caused the decline.
Why: Google documents deduplication; Ahrefs measures feature prevalence and correlation, not your expected traffic outcome.
Test: Pass when reporting compares verified feature appearances with clicks and qualified visits. Fail when it counts the snippet and a duplicate first-page listing as simultaneous exposure.
Measuring what you earned
Keep separate records for eligibility, observed appearances and resulting visits. Search Console can establish indexing status and report performance, but an average position near one does not identify a featured snippet.
Google’s performance counting documentation says a featured snippet typically occupies position 1. An AI Overview occupies one shared position, assigned to all its links. Expanding People Also Ask is not a click; links inside an expanded answer can receive impressions.
Google’s AI optimization guide tells site owners to use Search Console’s Generative AI performance report to monitor visibility in AI features. As of 31 August 2026, the report has rolled out to websites worldwide. It shows AI Overview and AI Mode impressions over time, with page, device and country breakdowns. It does not identify citations of specific passages, and AI feature traffic also remains included under the Web search type.
Record the query, engine, observation time, cited URL and visible passage for sampled appearances. Keep those observations separate from referral sessions: a visible citation establishes exposure, while a recorded visit establishes a click that reached your site.
Bing’s controls govern a separate ecosystem, so Google reports cannot validate Copilot visibility. For ChatGPT, check OAI-SearchBot access separately from training permission and compare observed answers with referral logs; access logs prove retrieval, not citation. Apply the same distinction to PerplexityBot requests.
A citation can answer the buyer’s question without sending a visit. Use the traffic and business-outcome checks for AI answers to connect visibility with qualified demand, then annotate content changes and verified appearances against the pages and queries that produce that demand.
Put your next steps in order
Build on what you have just read with a free ebook on semantic SEO and AI visibility. Get worked examples, decision rules, checklists and a do-this-week action in every chapter, plus a six-week plan for your site.



