How this was measured
On 2026-09-25 we fetched https://<domain>/ for 78 hand-picked domains spanning news, commerce, SaaS, developer tooling, social, education, finance, government and SEO. 62 returned a usable page: 41 over a direct connection and 21 over a local proxy, because a number of these domains are unreachable from our network without one.
Three of the 62 were not homepages and were dropped. khanacademy.org answered 200 with a 3 KB Cloudflare interstitial whose entire title is "Client Challenge". paypal.com returned an 8 KB JavaScript shell with no <title> at all. archive.org returned a 1.9 KB shell. Counting them would have manufactured findings that do not exist — a 15-character title, two missing titles — so they are excluded and listed in the script's output instead.
One bug is worth naming, because it changed the headline. Our first pass read python.org as having no <title>. It has one. The site returns gzip unconditionally, and our fetch was not decompressing, so we were parsing gzip bytes as HTML. Once the fetch decompressed, the title appeared, and the count of homepages without a title went from two to zero. A measurement pipeline that fails quietly produces confident nonsense.
For each page we took the first <title> and the first <meta name="description">, decoded HTML entities, collapsed runs of whitespace and measured in Unicode code points. Skipping the decode step inflates every number: & counts as five characters instead of one.
Finding 1: every homepage has a title, and the median is 35 characters
59 of 59. Not one homepage in the sample ships without a <title>. The distribution is narrow at the bottom and long at the top: minimum 3 characters, 25th percentile 22, median 35, 75th percentile 54, 90th percentile 61, maximum 116, mean 36.9.
The familiar rule — keep your title under 60 characters — is true and almost empty. Only 7 of 59 pages (12%) exceed 60, and just 2 exceed 70. The real behaviour is at the opposite end: 21 of 59 (36%) are under 30 characters. Half the sample writes a title shorter than 35 characters.
If you have been treating 60 as a target to hit, you are aiming at a number almost nobody in this sample reaches.
Finding 2: the shorter the title, the bigger the brand
The shortest titles belong to the most famous names in the sample: w3.org (3 characters, "W3C"), forbes.com (6, "Forbes"), tumblr.com (6), twitch.tv (6), youtube.com (7), pinterest.com (9), wikipedia.org (9), wikimedia.org (9), kubernetes.io (10), developer.mozilla.org (12, "MDN Web Docs").
The longest belong to sites that still have to explain themselves: bbc.com at 116 characters, airbnb.com at 74, atlassian.com at 70, wise.com at 68, shopify.com at 66, github.com at 61.
That is not a style difference, it is an equity budget. "Forbes" is a complete title because the reader already knows what Forbes is; the word carries the meaning the other 54 characters would otherwise have to. A site nobody has heard of that ships its name alone has said nothing at all. The 60 characters are not a ceiling to stay under — they are space you have to fill until your name does the work by itself.
The practical corollary: copying the title format of a famous site means copying a luxury you have not earned yet.
Finding 3: nearly everyone writes a meta description, so that advice is stale
57 of 59 homepages (97%) ship a description. Only two do not: linkedin.com and cdc.gov.
This contradicts a piece of advice still repeated constantly — that most sites forget the description and let Google invent one. On this sample that is simply false. What the last decade of snippet rewriting actually did was make the tag universal rather than optional. The omission rate here is 3%, not the 50% the folklore implies.
Two caveats. linkedin.com served a 121 KB JavaScript shell and cdc.gov a 60 KB document, so a description injected after load would be invisible to this method; both are recorded as "none in the served HTML", which is also what a crawler parsing the initial document sees. And writing a description does not mean Google uses it — this measurement is only about what sites ship, not about which sentence ends up in the snippet.
Finding 4: descriptions cluster at 114 to 150, and the 160 guideline holds up
The length distribution: minimum 45, 25th percentile 114, median 131, 75th percentile 150, 90th percentile 184, maximum 559, mean 141.
That interquartile range — 114 to 150 — is remarkably tight for a number nobody enforces. Practitioners have converged on a working band just under the display limit without being told to.
12 of the 57 (21%) run past 160 characters. The extremes are instructive: kubernetes.io ships 559 characters, roughly three and a half times what will ever be shown; stanford.edu ships 364; trello.com ships 195. At the other end, four descriptions fall under 70 characters, leaving more than half of the available space empty.
The practical read: 120 to 155 characters is where this sample lives, and it is also where truncation is least likely.
Finding 5: pipe is the most common separator, but no separator is more common
Counting only separators with whitespace on both sides, so that the hyphen inside rust-lang is not mistaken for one: pipe appears in 16 titles (27%), hyphen in 11 (19%), em dash in 2, and colon, middle dot and slash once each.
But 27 of 59 (46%) use no separator at all. The pipe is the plurality, not the majority; nearly half the sample writes a name or a sentence and stops.
A separator is a formatting decision with no ranking consequence. What it does signal is how many ideas you are concatenating. One separator means two ideas, and in most of the titles above the second one is the weaker — the tagline, the region, the boilerplate. The titles with no separator have a single idea, and a single idea is usually the better title.
Finding 6: your title is not one thing
We fetched from a Chinese IP address, and several sites served a regional variant rather than the default. linkedin.com returned 领英企业服务. stripe.com returned "Stripe | 金融基础设施，托举营收增长". wise.com returned "Wise: The international account | Money without borders | Wise China". bloomberg.com returned "Bloomberg Europe".
None of those are mistakes. But they mean a title audit performed from one location gives you one slice of a title that changes with geography, and a crawler dispatched from a different location reads a different string. The same applies to the description.
If you localise titles per region, decide it deliberately and keep it consistent with your hreflang annotations and your canonical choice. If you do not intend to localise, check that your edge is not doing it for you — the way to find out is to fetch your own homepage from more than one country.
What to do with this
Write 50 to 60 characters, not 3. The median homepage title in this sample is 35 characters and 36% are under 30 — but those are household names. Unless yours is one, use the space.
Always write a description. 97% of the sample does. It is table stakes now, not an optional extra, and omitting it hands the sentence a searcher reads entirely to Google.
Target 120 to 155 characters. That is the measured interquartile band, and it sits just inside where truncation begins.
Do not copy a famous site's title. Its brevity is backed by recognition you have not earned yet.
Prefer one idea per title. 46% of the sample uses no separator. When you do need one, pipe is the safe default at 27% — but check whether the second half is pulling its weight.
Check what your edge actually serves. Fetch your homepage from two or three regions before concluding that your title is what you wrote.
To see the truncation before you ship it, the SERP preview tool on this site renders a title and description at Google's real desktop and mobile widths. The meta tag generator emits the title, description and canonical from a single input, so the values cannot drift apart from each other.
Limitations
59 homepages from 78 hand-picked, well-known domains — representative of famous sites, not of the web. A random sample would look different in both directions, and probably longer: small sites write longer titles than W3C does.
16 of the 78 domains never produced a readable page — 403s, bot checks, timeouts. That loss is not random; it skews toward large consumer and financial sites that fight crawlers, so the sample is biased toward sites that serve their homepage to anyone who asks.
We read the initial HTML only. A title or description injected by JavaScript after load is invisible here, and one of the two missing descriptions may be exactly that.
One fetch, from one IP address, on one day. Finding 6 is the direct consequence of that limitation rather than a separate discovery.
The brand-position analysis we planned did not survive checking and is not reported. Judging brand placement from the domain's first label fails on multi-word brands and on brands spelled differently from their domain: 9 of the 10 titles flagged as containing no brand did contain one — "the Guardian", "Node.js", "Rust", "Screaming Frog", "Search Engine Journal", "Square". We dropped the metric rather than publish a number we could not stand behind.
Reproduce it
The script and the shared domain list ship with this site. node scripts/survey-serp-snippet.mjs fetches and stores the raw HTML for every domain; --report re-derives every number from that stored HTML without touching the network; --sample=N prints raw title and description lines for manual checking; and --outliers lists every title over 60 characters, every description over 160, and every response excluded as a non-homepage. Our robots.txt, hreflang and canonical surveys cover the same domain list from different angles.