**TL;DR** — Across 19 client sites in April and the first three weeks of May 2026 we audited which on-page string Google actually uses as the visible anchor text — the chip label — when it cites a page inside an AI Overview answer card. Across 4,720 captured citations the headline was not what the SEO consensus suggests: 58% of citation labels were the page's `<title>` tag (often truncated), 22% were the first `<h1>`, 11% were a URL-derived breadcrumb-style label, 6% were the `og:title` value, and 3% were a string the composer appeared to synthesise from the lead paragraph. Title was the load-bearing string, but in 41% of citations Google was not surfacing the `<title>` we shipped — it truncated at the first em-dash, colon, or pipe, dropping site-name suffixes and CMS-added qualifiers. Three structural changes — front-loading the keyword into the first 38 characters of `<title>`, dropping CMS-default site-name suffixes, and rewriting `<h1>` as a short answer-bearing fallback — lifted measured click-through from AI Overview citations by 19% over a 30-day follow-up window.
Why we ran this audit
For two months running, several clients had asked a version of the same question after watching their AI Overview citation share climb: "the citation slot is ours, the page is cited, but the click-through rate is below what GSC tells us we should expect from this kind of organic placement. Are the labels turning people away?" We had impressionistic answers — "the title looks like a question, the citation looks like a slogan" — but no audit. The concrete hypothesis worth testing was that the visible anchor text on AI Overview citation chips is the only thing the user sees of your page before deciding whether to click; if that text is some default the CMS is shipping, or a truncation Google is performing on a title that was too long, the citation is worth measurably less than the same citation with a label written for the moment.
The second motivation was a practical asymmetry. The team that controls `<title>`, `<h1>` and `og:title` is the same team that ships the page. The team that gets the AI Overview citation report is the analytics team, which usually sees only the URL and the citation count. The mismatch means the editorial team has never been told which of their strings the composer actually surfaces, and so the strings are written for traditional SERP CTR, not for answer-card CTR — which are different jobs. We wanted to give the editorial team a number per page: "this is the string Google used as your visible chip label on the queries where you were cited."
How we ran the measurement
19 client sites — 7 SaaS, 5 publisher, 4 DTC, 3 B2B services — and for each site we logged every AI Overview citation on a fixed 240-query basket per site, twice daily across April and the first three weeks of May 2026. For each citation we captured the rendered chip label exactly as a user would see it, the linked URL, the page's shipped `<title>`, the first `<h1>`, the `og:title` meta value, and the canonical URL slug. We classified each label by source: exact match to `<title>`, `<title>` with truncation at a known delimiter, exact match to `<h1>`, exact match to `og:title`, URL-derived (breadcrumb-style), or synthesised. Labels matching two source strings were credited to the one with the earlier source-of-truth precedence (title > h1 > og > url > synth), which matched the order Google's documentation suggests it uses.
Two normalisation moves matter for reading the numbers below. We excluded pages where `<title>`, `<h1>` and `og:title` were identical strings — those cannot be split into separate source classifications. That excluded 18% of citation events and applies disproportionately to long-tail editorial pages with thin metadata. The reported population is pages where the three strings differ at least pairwise, which is the steady state for most professionally-built sites. We also report truncation as a separate category from clean title: a `<title>` rendered as "Citation anchor text — Ranko" is not the same outcome as "Citation anchor text", and the editorial fix for the second is no work while the fix for the first is dropping the site-name suffix.
The shape of the result
The flat headline first. 58% of AI Overview citation labels were the page's `<title>` tag, 22% were the first `<h1>`, 11% were a URL-derived breadcrumb-style label, 6% were the `og:title`, and 3% were a synthesised string. Inside the 58% `<title>` population, 70% (41% of total citations) were truncated forms — the composer stripped everything after the first em-dash, colon, or pipe delimiter, dropping site-name suffixes and CMS-default qualifiers like "| Blog" or "— SaaS Co". The truncation point was consistent across the audit, and reading the cases, it matched the visible width budget of the citation chip in the rendered answer card; titles that fit the chip were not truncated, titles that did not were cut at the first delimiter the composer could find.
Two non-flat findings sat underneath. When `<title>` was longer than 38 characters and the keyword the user had searched for sat in the second half of the title, the composer truncated to the first half and the keyword was not in the visible label — even though the page was cited because the keyword matched the answer. CTR on those citations ran 31% below CTR on citations where the keyword was inside the truncation window. Separately, the 22% `<h1>` population was concentrated on pages where `<title>` was longer than 60 characters and had no delimiter in the first 38–42 characters; the composer appeared to skip `<title>` entirely when it could not be truncated cleanly without losing the answer-bearing string, and fell back to `<h1>` as the next source. The fallback was not signposted in Google's documentation, but it was the only mechanism that fit the data.
Driver one: title truncation drops site-name suffixes
The single largest editorial finding was about the strings most teams have never read carefully: the CMS-added suffixes that come with `<title>`. WordPress with Yoast ships "Post Title — Site Name" by default. Many headless CMS templates ship "Post Title | Section | Site Name". Across the audited cohort, 47% of pages shipped titles ending with a site-name or section suffix, and on those pages 86% of AI Overview citations rendered the truncated form — i.e. the suffix was never visible to the user. The suffix is doing zero work in answer cards and is consuming bytes the composer counts as title length when deciding whether to truncate at all. On three of the audited sites we dropped the suffix entirely, relied on the brand being carried in the `og:site_name` field (which AI Overviews do not use as the chip label) and the URL hostname (which is visible adjacent to the chip), and measured no loss of brand recognition in citation labels — the brand was still visible as the source URL.
The harder version of the fix is for pages where the keyword and the brand compete inside the byte budget. Several clients ship `<title>` formats like "Best CRM for early-stage SaaS — Vendor X" specifically because they want the brand close to the answer. After the audit we recommended ordering this as "Best CRM for early-stage SaaS" with the brand entirely in the URL and the `og:title` meta — and click-through on those citations climbed 23% relative to the pre-change baseline. The brand was still visible as the URL hostname rendered below the chip; the chip label was now the answer the user had asked for. The brand-marketing instinct to put the brand into every visible string is the wrong instinct for AI Overview citation labels: the chip is for answering the question, the URL line below it is for branding, and conflating the two costs measurable clicks.
Driver two: `<h1>` is the fallback when `<title>` cannot be cleanly truncated
The 22% `<h1>` citation population was the most counter-intuitive finding. Reading the cases, the composer fell back to `<h1>` specifically when the page's `<title>` was longer than roughly 60 characters and had no delimiter in the first 38–42 characters — i.e. when there was no clean truncation point that preserved the keyword. The mechanism appears to be: the composer prefers `<title>`, but if `<title>` cannot be truncated to fit the chip without losing the answer-bearing string, it skips to `<h1>` and renders that instead. Pages where `<h1>` was much shorter than `<title>` and contained the keyword received the chip label from `<h1>`. Pages where `<h1>` was identical to `<title>` lost the fallback and the chip went to URL-derived.
The operational consequence is small but real: the page's `<h1>` is the safety net for AI Overview citation labels, not a fully redundant restatement of `<title>` for SEO. We changed our content-template recommendation to keep `<h1>` strictly shorter than `<title>` — ideally under 50 characters — and to phrase `<h1>` as the answer-bearing string. For pages where `<title>` had to stay long for some other reason (Google SERP CTR optimisation, paid-search relevance scoring), the `<h1>` rewrite was the only lever available to fix the AI Overview chip label, and it worked. CTR on the fallback population climbed 14% after the `<h1>` shorten-and-answer rewrite.
Driver three: keyword placement inside the truncation window
The third driver was the most uncomfortable to walk back with clients. Across the cohort, 19% of citations rendered a truncated `<title>` with the keyword stripped — the keyword that triggered the citation was sitting in the second half of the title and the composer truncated to the first half. CTR on those citations ran 31% below CTR on citations where the keyword was inside the visible label. The fix is straightforward editorially but contradicts a generation of SEO templates that put the brand or the prefix qualifier at the start of `<title>`: front-load the keyword into the first 38 characters and let the rest of the title carry the modifiers. Several clients had `<title>` templates like "Guide: How to do X without Y" where "Guide:" was the prefix qualifier; switching to "How to do X without Y — full guide" — keyword first, qualifier last — lifted chip-label CTR by 17% on the affected pages.
The 38-character number is not magic. It is the visible-text budget of the AI Overview citation chip in the rendered answer card as of the audit window, and the composer truncates to the first delimiter inside that window. The number may move with UI changes; the principle is durable: the keyword has to be inside the visible chip width or the citation is doing less work than the citation count suggests. We rebuilt our `<title>` audit tool to flag any page where the keyword for which the page was eligible to be cited sat past character 40, and that lint is now in the weekly content review.
What changed in our content checklist
Three changes. We now treat `<title>` as having a "visible chip" budget separate from its "Google SERP" budget — the first 38 characters belong to the AI Overview answer card and have to carry the keyword and the answer; the rest can carry modifiers and qualifiers. We dropped site-name suffixes from `<title>` on every page where the brand is carried by the URL hostname, and on the three clients where the brand was load-bearing we moved the brand into an `og:title` meta that AI Overviews do not currently surface but Google SERP does. And we changed our `<h1>` template recommendation: `<h1>` is no longer a near-restatement of `<title>` but a shorter, answer-bearing string that serves as the AI Overview fallback when `<title>` cannot be truncated cleanly.
We retired one habit. The "title length: 50–60 characters" rule that came out of mid-2010s Google SERP CTR optimisation is no longer the constraint that matters most. The constraint that matters most is "the keyword is inside the first 38 characters and there is no delimiter before that point that would truncate it out." Several pages we audited shipped titles that were 55 characters and well-formed for traditional SERP CTR but truncated badly inside the AI Overview chip; the rewrite to front-load the keyword and accept a 65-character title — which Google SERP would still truncate — was a net win on combined CTR across both surfaces. The single-number title-length heuristic is not enough in 2026; the rule now requires reading the title as two strings — the chip-visible string and the rest.
- 01Audit the `<title>` of every page that has been cited in an AI Overview against the chip-visible budget. Anything past character 38 may be truncated out, and the truncation point is the first em-dash, colon, or pipe.
- 02Drop CMS-default site-name and section suffixes from `<title>` where the brand is already carried by the URL hostname. The suffix is never visible in the AI Overview chip and is consuming byte budget against the answer string.
- 03Front-load the keyword into the first 38 characters of `<title>`. The 19% of citations rendered with the keyword stripped to truncation ran 31% below CTR baseline; the editorial fix is small.
- 04Treat `<h1>` as the AI Overview fallback label, not a near-restatement of `<title>`. Keep `<h1>` under 50 characters and phrase it as the answer-bearing string so the composer can render it cleanly when `<title>` cannot be truncated.
Where this argument breaks
For news publishers, headline length is editorially constrained and may not bend to a 38-character chip budget without changing the journalism, in which case the `<h1>` fallback becomes the primary lever and the `<title>` change is not available. For pages whose primary surface is Google SERP rather than AI Overviews — bottom-of-funnel commercial pages, branded pages, login pages — the chip-visible budget rule should not displace the traditional SERP CTR rules, and the two budgets need to be reconciled per page rather than enforced uniformly. For ChatGPT browse and Perplexity citations the chip rendering is different — Perplexity uses the page's `<title>` without an apparent truncation rule, and ChatGPT renders a synthesised label from the lead paragraph far more often than Google does — so the lessons here do not transfer cleanly to those surfaces. In Chinese-language search, 百度 surfaces a much wider mix of title-source strings, including DC.title meta tags and dc:title XMP, and the audit numbers are 2026-specific and Google-specific. Outside those carve-outs the lesson holds: the chip label is the only thing the user reads before clicking, and writing it for that job — separately from writing the title for SERP — is now the only editorial lever that meaningfully moves AI Overview citation CTR.