Methodology
We started with 240 real Google searches across nine industries: legal software, careers, personal finance, plumbing, cybersecurity, supplements, plastic surgery, home renovation and crypto.
For each search we recorded which pages Google's AI Overview cited, on September 11, 2026. Then on September 14 we asked Perplexity, Gemini and Claude the same question and recorded which pages each one cited. For Claude and Gemini we also recorded the searches they ran before answering, and that turned out to matter a lot.
We compared the cited pages against pages that ranked in Google's top 30 for the same search but weren't cited. That's 3,486 distinct pages, showing up in 4,033 page and search combinations. For every page we read the title tag, the H1, and the published and updated dates in the page's code, and we scored how closely the title matches the search.
How to read the numbers
"In 100." Take 100 pages with a feature and 100 without, competing for the same searches. How many of each got cited?
Lift. How many more citations the pages with the feature got. So 19 cited in 100 against 6 in 100 is +234%. We calculate lift from the unrounded rates, so it won't always match dividing the rounded numbers.
Within the same search. Some searches are just easier to get cited in than others. Every comparison here is between pages competing for the same search, so the gap is about the page and not the search.
A quick warning about big percentages. Lift gets huge when the starting number is small. Going from 2 cited in 100 to 17 is "+750%", but it's still 15 extra pages per 100. That's why every lift in this report sits next to its "in 100" numbers.
How we checked each pattern held up
A pattern can look strong on 240 searches and still be luck. So we split the searches in half at random, checked whether the pattern showed up on both halves, and did that 30 to 40 times per pattern.
| Rating | What it means |
|---|---|
| Reliable | Showed up in 95% or more of re-tests |
| Likely | Showed up in 70% to 94% of re-tests |
| Mixed | Showed up in 40% to 69% of re-tests, on Google or Perplexity |
| Not supported | Showed up in fewer than 40% of re-tests, on Google or Perplexity |
| Unconfirmed | Claude or Gemini, below 70%. Too little data to confirm it or rule it out |
Claude and Gemini get their own rating because we have a lot less data on them. They only search the web for some questions, so Gemini answered 129 of our searches with citations and Claude answered 102. A pattern that fails to show up on that little data is weak evidence either way. A pattern that does show up on it, like the year in the title, is strong evidence.
These ratings describe how consistent a pattern is. They don't prove that changing your title causes a citation, and we come back to that in the limitations.
Finding 1: A year in the title gets more citations on Claude, Gemini and Perplexity
Within the same search, across pages ranking in Google's top 30:
| AI engine | Cited, year in title | Cited, no year | Lift | Held up? |
|---|---|---|---|---|
| Claude | 19 in 100 | 6 in 100 | +234% | Reliable |
| Gemini | 12 in 100 | 5 in 100 | +138% | Likely |
| Perplexity | 23 in 100 | 14 in 100 | +65% | Reliable |
| Google AI Overviews | 21 in 100 | 17 in 100 | +22% | Not supported |
This is unusually strong for Claude and Gemini. Patterns usually struggle to show up consistently on that little data, and the year still held up in 100% of re-tests on Claude and 88% on Gemini.
It also holds after accounting for the obvious explanations. On Claude, Gemini and Perplexity the lift stays as large or gets larger once we account for Google rank, page length, site authority, the page's headings and how well its content matches the search.
It isn't just that pages with a year are newer
That was the first thing we wanted to rule out. Pages with a year in the title are more likely to be recent, so maybe the year is just standing in for freshness.
So we looked only at pages last updated in 2025 or later. A year in the title still went with more citations:
| AI engine | Year in title | No year | Lift |
|---|---|---|---|
| Claude | 20 in 100 | 7 in 100 | +171% |
| Gemini | 13 in 100 | 5 in 100 | +158% |
| Perplexity | 21 in 100 | 16 in 100 | +30% |
| 22 in 100 | 21 in 100 | +3% |
And the year should be current. Among titles that carry a year, titles with 2026 got 214% more citations on Claude (24 against 8 in 100) and 169% more on Perplexity (24 against 9) than titles with an older year. Out of date years are rare though, 2 or 3 in every 100 titles, so we can't tell you whether an old year is worse than no year at all.
The current year comparison only covers the 14 in 100 titles that carry a year, on 39 to 60 searches per engine, and we didn't re-test it. So treat it as a direction.
Why it works: the AI adds the year to its own search
This is the part that actually explains it. Before answering, Claude and Gemini run their own web searches, and our data records what they searched for.
Most of the time, they add a year. Asked for "best resume builder", Claude searched for "best resume builder 2026". So a page with 2026 in its title matches what the AI actually typed, even though the person never said a year.
| Claude | Gemini | |
|---|---|---|
| Searches where the AI added a year | 68 in 100 | 60 in 100 |
| Which year | 2026, almost always | 2026 most often, but 2024 and 2025 a lot too |
| Searched with a year: year in title vs none | 21 vs 2 in 100, about 10x | 15 vs 4 in 100, about 4x |
| Searched without a year: year in title vs none | 12 vs 19 in 100, 17 searches | 0 vs 13 in 100, 15 searches |
When the AI put a year in its search, pages with a year in the title dominated. When it didn't, the advantage went away. But only 15 to 17 searches fall in that second group, so it's a small sample.
There's one catch here. The AI tends to add a year on the kinds of searches where a year is natural anyway, like "best of" lists and pricing. So this overlaps with the type of search. It's still the clearest explanation we have.
Gemini also searches with out of date years sometimes. It added 2024 to its searches 80 times in this September 2026 snapshot, which is another reason we can't say whether an old year in your title hurts on Gemini.
Perplexity doesn't show us its searches, so we can't check whether the same thing is happening there.
Where a year belongs
Most year in title pages in our data are "best of" and pricing pages, and that's where the lift is clearest. Within the same search, across Google's top 30:
| Search type | Perplexity | Gemini | Claude | |
|---|---|---|---|---|
| "Best of" searches | 21 vs 8 (+160%) | 16 vs 2 (about 9x) | 17 vs 2 (about 10x) | 16 vs 12 (+29%) |
| Cost and pricing searches | 19 vs 9 (+100%) | 13 vs 6 (+114%) | 20 vs 10 (+92%) | 19 vs 17 (+10%) |
Cited in 100, year in title vs no year. Each cell rests on 18 to 37 searches, so treat these as a guide to where to start rather than a forecast. The "about 9x" and "about 10x" start from 2 in 100 and will probably shrink with more data.
How-to content went the other way. On how-to searches, the few pages with a year in the title got 54% fewer citations on Perplexity, 11 against 23 in 100 across 16 searches. A year on evergreen content might just look out of place. There were too few pages to be sure.
Finding 2: Google cares whether the page was actually updated
Google behaves differently from the AI assistants. The year in the title did nothing reliable on Google, but real freshness did.
About 70 in every 100 pages in our data publish a "last updated" or "published" date in their code. Among those pages, within the same search:
| Where | Updated 2025 or later | Older | Lift | Held up? |
|---|---|---|---|---|
| Google, top 30 | 21 in 100 | 7 in 100 | +220% | Reliable |
| Google, first page only | 54 in 100 | 29 in 100 | +89% | |
| Perplexity, top 30 | 19 in 100 | 11 in 100 | +82% | Not supported |
| Claude, top 30 | 13 in 100 | 5 in 100 | +162% | Unconfirmed |
| Gemini, top 30 | 10 in 100 | 7 in 100 | +49% | Unconfirmed |
On Google, the lift survives accounting for rank and site authority, where it's still about +105%, and it showed up in every single re-test. On the other three, most of it disappears once rank and authority are accounted for, down to +15% to +40%, and it didn't show up consistently.
So you get a pretty clean split:
- Google responds to whether the page is genuinely recent. Once that's known, a year in the title adds nothing.
- Claude, Gemini and Perplexity respond to the year in the title. Whether the page was genuinely updated didn't show up consistently for them.
The practical answer is to do both. Actually update the page, make sure the updated date is published in the page's code, and put the current year in the title.
Changing the year without updating the page gets you nothing on Google. And if an assistant cites a stale page because of a new year in the title, you're sending people to out of date information, which isn't a great look for your client.
Finding 3: A title that contains the search helps on Perplexity
On Google's first page, within the same search:
| AI engine | Title contains the exact search | Title doesn't | Lift | Held up? |
|---|---|---|---|---|
| Perplexity | 38 in 100 | 32 in 100 | +20% | Reliable |
| Claude | 41 in 100 | 24 in 100 | +71% | Unconfirmed |
| Google AI Overviews | 44 in 100 | 38 in 100 | +15% | Not supported |
| Gemini | 16 in 100 | 16 in 100 | +5% | Unconfirmed |
A looser version, where the title contains every word of the search in any order, gives you the same picture. It's +15% on Perplexity (Reliable), +76% on Claude (Likely) and +13% on Google (Not supported).
Claude's number looks big, but it rests on 60 searches and the exact version didn't hold up when we re-tested it. So we'd lean on the Perplexity result.
On Google, the rest of the page already does this job
Once we account for whether the page has a heading that asks the searcher's question, and content that matches the search, a title containing every search word adds only +8% on Google, and not consistently.
On Perplexity it still adds +44% on its own. On Claude it adds +53%, with less data behind it.
Matching the meaning, and not just the words, was Mixed. We scored how closely each title means the same thing as the search, even without the exact words. Titles in the top quarter for that score got +20% on Google's first page (47 against 39 in 100) and +33% on Perplexity (37 against 28). But that only showed up in 60% of re-tests on Google and 67% on Perplexity.
The main heading on the page, the H1, behaves a lot like the title. An H1 containing the exact search got +11% on Perplexity's first page, which was Reliable, and +25% on Google's, which wasn't.
Finding 4: The title tactics that made no difference
These are all pretty common title "best practices". None of them showed a consistent lift on any engine.
| Title tactic | What we found | Held up? |
|---|---|---|
| Short titles, under 40 characters | Looked like +21% to +35% in raw counts, but shrank to under 8% on Google and Perplexity once rank, page length and site authority were accounted for | Not supported on any engine |
| Long titles, over 65 characters | Between -19% and +16% across engines, with no pattern | Not supported on any engine |
| "| Brand Name" or "- Brand Name" suffix | +7% on Google, +3% on Perplexity, -4% on Gemini, -35% on Claude, none of them consistent | Not supported on any engine |
| Title written as a question | +24% on Google's first page, but it held up in only 10% of re-tests. -40% on Gemini and -25% on Claude | Not supported |
| A list number, like "7 best" | Between -16% and +31% across engines once years are separated out | Not supported on any engine |
We got the list number wrong the first time
Our first pass said "a number in the title" went with more citations on Perplexity and Claude. That was actually the year.
Years are numbers, so "Best CRMs 2026" was being counted as a numbered title. Once we separated years out, list numbers showed nothing at all. We're leaving that in because it's exactly the kind of mistake that ends up in a best practices list.
Put the question in a heading instead
A question-style title didn't help. But in the same dataset, a heading on the page that asks the searcher's question went with 45% more citations on Google (48 against 33 in 100 on the first page), and it held up in every re-test on Google.
So the question belongs on the page. The title doesn't need to be one.
What to do with this
- Put the current year in the title on "best of", pricing, review and ranking pages. Those pages got +234% on Claude, +138% on Gemini and +65% on Perplexity. Leave it off evergreen how-to and definition content.
- Actually update those pages, and publish the updated date in the page's code. Genuinely updated pages got +220% on Google. A new year on an old page gets you nothing there.
- Keep the year current, and refresh the page when you change it. A current year beat an older year by +214% on Claude and +169% on Perplexity.
- Include the words of the target search in the title, mainly for Perplexity, where it's worth +15% to +20%. On Google, your headings and content already do this job.
- Stop spending time on title length, brand suffixes and question titles for AI search. We found no consistent lift for any of them. They might still matter for clicks from normal search results, which we didn't measure.
If you want to know whether this works on your own pages, the most reliable way is to test it. Take a group of similar "best of" pages and genuinely update all of them. Add the current year to the title on half, then compare citations on Perplexity and Claude a month later.
Every pattern, and whether it held up
| Pattern | Perplexity | Gemini | Claude | |
|---|---|---|---|---|
| Year in title | Not supported | Reliable | Likely | Reliable |
| Page updated 2025 or later | Reliable | Not supported | Unconfirmed | Unconfirmed |
| Title contains every search word | Not supported | Reliable | Unconfirmed | Likely |
| Title contains the exact search | Not supported | Reliable | Unconfirmed | Unconfirmed |
| Title means the same as the search | Mixed | Mixed | Unconfirmed | Unconfirmed |
| H1 contains the exact search | Not supported | Reliable | Unconfirmed | Unconfirmed |
| Title written as a question | Not supported | Not supported | Unconfirmed | Unconfirmed |
| List number, years removed | Not supported | Not supported | Unconfirmed | Unconfirmed |
| Title length | Not supported | Not supported | Unconfirmed | Unconfirmed |
| Brand name suffix | Not supported | Not supported | Unconfirmed | Unconfirmed |
The raw re-test rates behind those ratings:
| Pattern | Perplexity | Gemini | Claude | Re-tests | |
|---|---|---|---|---|---|
| Any year in title | 0% | 98% | 88% | 100% | 40 |
| 2026 in title | 0% | 100% | 90% | 100% | 40 |
| Updated 2025 or later | 100% | 0% | 0% | 0% | 40 |
| Title contains every search word | 0% | 100% | 0% | 73% | 30 |
| Title contains the exact search | 0% | 100% | 0% | 40% | 30 |
| Title relevance score | 60% | 67% | 0% | 3% | 30 |
| H1 contains the exact search | 3% | 100% | 0% | 27% | 30 |
| Title is a question | 10% | 0% | 0% | 0% | 30 |
| List number, years removed | 0% | 0% | 0% | 2% | 40 |
| Title under 40 characters | 0% | 0% | 0% | 0% | 40 |
| Title over 65 characters | 0% | 0% | 0% | 0% | 40 |
| Brand separator | 0% | 0% | 0% | 0% | 40 |
Limitations
These are patterns, not proven cause and effect. Pages with a year in the title might differ in ways we didn't fully measure, and the biggest candidate is page type, like "best of" lists that get refreshed every year. Comparing within the same search and checking only recently updated pages rules out the obvious explanations, but not all of them.
Timing matters. We captured these answers in September 2026, when 2026 was well established as the current year. Early in a new year the engines might still search for the previous one, and Gemini was already using older years in a lot of its searches. We don't know how quickly they switch over in January.
Some checks came after we saw the first results. The title features were fixed before any results came in. But several follow-ups were added afterwards to explain what we saw: the list number split, which year, the recently updated pages check and the AI's own searches. Treat those as explanation rather than independent confirmation.
Freshness depends on the date a page publishes. 30 in every 100 pages publish no date we could read, so they're left out of Finding 2. We used the page's stated last updated date, or its published date where there was no update date.
This covers part of what each engine cites. We only compared pages that ranked in Google's top 30. That covers 38% of the pages Google's AI Overviews cited in our searches, 35% for Claude, 28% for Perplexity and 18% for Gemini. The rest were pages outside Google's top 30, and this study doesn't describe them.
The engines answered different mixes of searches. Claude and Gemini only searched the web for some questions, and more of those were "best of" and pricing searches. So don't read the lifts as a ranking of which engine cares most.
Clicks aren't measured. This is about getting cited in AI answers, not click-through from normal search results.
US, English, desktop. Results could look different elsewhere.
ChatGPT isn't included. The data source we used returned too few cited pages to measure, so this says nothing about ChatGPT either way.
Technical appendix
Method. The "in 100" figures are within-search paired rates. For each search with pages on both sides, we took the citation rate with the feature and without it, then averaged across searches. Lift is the ratio of those two averages minus 1, from unrounded rates.
The ratings come from conditional logistic regression grouped by search, controlling for Google rank, log word count, Domain Rating and URL Rating, run on random half-splits. A pattern counts as showing up when it's significant in the same direction on both halves. Each engine is modelled only on searches it answered with web citations: Google 240, Perplexity 240, Gemini 129, Claude 102.
Feature definitions.
- Year in title: any year from 1980 to 2039 in the title tag. That's 14.0% of pages, split into 2026 at 11.6%, 2025 at 1.8%, and 2024 or earlier at 0.6%.
- List number: any digit left after years are removed, 18.9% of pages.
- Exact search: the full query string appears in the title. Every word: all meaningful query words appear, in any order.
- Relevance score: TAS-B (msmarco-distilbert-base-tas-b) similarity between the title and the query.
- Updated 2025 or later: dateModified of 2025 or later, falling back to datePublished, article published and modified times, or a time element where there's no modified date.
Year in title, adjusted lift. Converted from conditional logit odds ratios at each engine's base rate, all ranks. With the study controls only: Google +25%, Perplexity +109%, Gemini +180%, Claude +288%. Adding passage relevance and question headings as controls: Google +16%, Perplexity +94%, Gemini +171%, Claude +270%. Adding list number as a control: Perplexity +103%, Gemini +183%, Claude +278%. The headline lifts in this report use the simpler within-search rates (+65%, +138%, +234%), which are the more conservative of the two.
Year in title, first page only (within-search paired): Google 50 vs 35 across 97 searches, Perplexity 37 vs 26, Gemini 12 vs 3, Claude 41 vs 9. The Google first page gap doesn't survive the re-test (0% of splits at all ranks) or the recently updated pages check (22 vs 21, +3%).
Freshness, within-search. Dated 2025 or later vs older, all ranks: Google 21.4 vs 6.7 (+220%, 203 searches, 100% of re-tests, adjusted +105%), Perplexity 19.4 vs 10.6 (+82%, 0%, adjusted +15%), Gemini 10.0 vs 6.7 (+49%, 0%, adjusted +26%), Claude 13.5 vs 5.1 (+162%, 0%, adjusted +40%).
Fan-out queries. Taken from DataForSEO's fan_out_queries field. Claude: 164 queries across 102 searches, with a year on 69 searches (2026: 69 mentions, 2025: 1). Gemini: 457 queries across 129 searches, 123 of them with fan-out logged, with a year on 77 (2026: 138 mentions, 2024: 80, 2025: 48). Year in title lift when the engine searched with a year: Claude 21.1 vs 2.1 in 100 (+883%, 64 searches), Gemini 15.2 vs 3.7 (+312%, 70). Without a year: Claude 11.5 vs 19.2 (-40%, 17), Gemini 0.0 vs 12.6 (15). Perplexity Sonar returns no fan-out queries.
Why a low rate on Gemini and Claude is Unconfirmed rather than Not supported. On Google and Perplexity, a pattern that holds in 100% of re-tests on 240 searches holds in only 43% to 62% when we cut it down to 105 to 135 searches. So a low rate on the smaller engines is weak evidence either way, and a high rate on them, like the year in the title, is strong evidence.
Scope: 240 searches across 9 industries, 3,486 distinct pages, Google AI Overviews, Perplexity, Gemini and Claude (US, desktop, English), captured September 11 to 14, 2026. Google rankings, AI answers and on-page data from DataForSEO, site authority from Ahrefs.