# Twiddlers and page dates: how Google reshuffles results after ranking, and which date it trusts

> After the main ranking, Google runs results through twiddlers, small programs that boost, demote, filter and space them out, freshness included. What the 2018 document, the 2024 leak and RESONEO's 2026 catalogue show, which dates Google keeps for a page, and why a new date on unchanged text makes it unreliable rather than fresh.

_Source: https://seomatrix.ai/blog/twiddlers-freshness-dates/ · Updated: 2026-11-15_

---

Straight to the spoiler: after Google ranks results, more than a hundred small programs called **twiddlers** (by RESONEO's count) can still reshuffle them, and freshness has one of its own. Google keeps at least four dates per page plus a score of how far it trusts them. A new date on an unchanged page lowers that trust instead of buying freshness.

In early November RESONEO closed its six-part "Understanding Google" series with a piece on twiddlers. I read it with our research notes open in the next tab and one question in mind, the one I hear most about old content: can't we just bump the dates? Most CMSs will do it for a thousand posts in one click (and people do exactly that). So does a new date help, hurt, or do nothing? To answer, I had to go one layer down, to where Google moves results *after* it has decided how good they are.

## What a twiddler is

The shortest definition comes from notes of a January 2025 call with Nayak of Google, published by the US Justice Department in the antitrust case: "Twiddlers – re-rank a set of already selected results". RESONEO's version is the one I'd frame:

> Between ranking and display, twiddlers, small programs, rework the list of results: they boost one result, demote another, filter out a third or space out those from the same site. In 2018, an internal Google document counted more than 65 of them in production for web search alone.
> — RESONEO, [Twiddlers, November 2026](https://think.resoneo.com/twiddlers/)

The pipeline, as RESONEO reconstructs it from that 2018 "Twiddler Quick Start Guide" and the 2026 schemas, is short. Ascorer scores hundreds of results. **Predoc** twiddlers run over the whole list before snippets exist. A framework reconciles their advice and reorders. **Lazy** twiddlers then work on the head of the list with snippets in hand; if they push results out, the framework pulls in the next ones. Finally Tangram packs web, images and video into one page. Note the scope: a twiddler works inside one corpus, never across the whole results page.

A twiddler doesn't move anything itself. It recommends: `Boost`, `BoostAboveResult`, `filter`, `max_total` (cap results of one kind), `stride` (space them out) or annotating. "Each twiddler works alone, unaware of the others' decisions: it recommends, the framework decides."

Why so many? Economics. Per the 2018 document, a twiddler experiment needs a few Superroot jobs; the alternative is a section in Ascorer or rolling out 1,400 jobs. Cheap to launch, and RESONEO adds that "once launched, a twiddler is there to last". RankLab logs every pass, so Google can take any adjustment apart query by query. We can't, sadly!

My takeaway, and it's an interpretation, not a measurement: when positions jump while nothing changed on the site, a twiddler launch is probably a better suspect than you think, so compare the drop across a class of queries, not one page. Proof? No. Just a better first hypothesis than "the algorithm hates me".

## 120 twiddlers, and how much of that is just names

Here's the part most retellings skip. RESONEO published **only schema type names**, no fields. Descriptions like "expertise is assessed per page" are inferred from names such as `…DocFeatures` and from Google's public statements. The schemas' authenticity isn't confirmed, and RESONEO doesn't say where they came from. Weights and thresholds are unknown. Even the count, 120 twiddlers with 111 absent from both the 2018 document and the 2024 leak, is RESONEO's own.

What the 2024 leak does confirm, in Google's own field descriptions, is a handful of names: FreshnessTwiddler, SiteBoostTwiddler, WebImageBoostTwiddler, the Skin Tone Twiddler, plus a `hostAge` field "used in twiddler to sandbox fresh spam in serving time".

| Family (RESONEO, Nov 2026) | Count | Examples |
|---|---|---|
| AI and rankers | 6 | fan-out plan, Bard filter, relevance oracle, part of RankBrain |
| Quality | 9 | Expertise, Authenticity, Topic Authority, Web Health |
| Demotions | 12 | SiteBoost (UGC spam), site-level DMCA, crime demotion, SafeSearch |
| Freshness, repeat queries, personalization | 14 | FreshnessTwiddler v2, date score, view demotion, For You |
| SERP layout and verticals | 26 | Forums, Deals, Specialty merchants |
| Images | 20 | image boost, host diversity in the viewer |
| Video | 33 | YouTube clickbait, Shorts authenticity, dedup |

Three rows are worth a second look. The AI Mode fan-out plan "travels with the query into the second pass and sits among its ranking signals", so twiddlers know the planned searches (more in [our post on query fan-out](/blog/query-fan-out/)). Copyright now hits the site: past a certain volume of valid notices, the whole domain goes down, where in 2018 DMCAFilter removed single results. And forums, deals and specialty merchants have their own twiddlers, so a better-scored page can lose a shopping query to a thread. Annoying? Very!

> **A name in a schema is not a weight** — A twiddler's name tells you a mechanism exists in the code. It says nothing about its strength, when it fires, or whether it runs for your query. "There's an Expertise twiddler" is a fact about a codebase; "so expertise lifts rankings by X" is a guess. Same for patents: described is not switched on.

Also explicitly **not** confirmed: QualityBoost isn't called a twiddler anywhere, and the term QDF ("query deserves freshness") appears neither in the leak nor in RESONEO's work.

## Where freshness comes in

The freshness twiddler is the one I trust most here, because the 2024 leak names it outright: `datesInfo` is "Used in FreshnessTwiddler", and `semanticDateInfo` holds data "required by the freshness twiddlers". RESONEO's 2026 catalogue adds a second version and a separate **date score** whose parameters arrive with every query. Per RESONEO, the twiddler reads the page's dates and a content date Google estimates from the text, anchors and related pages. Which date fields exactly? Not shown: none of the sources tie it to `lastSignificantUpdate` or `latestFirstseenSec` specifically.

Freshness also seems to be weighed against the query, not in a vacuum. RESONEO's example: for knee pain while running, pages from 2024 and 2025 sit in the top next to October 2026. In their snapshot of "mercury retrograde 2026" on 28 October 2026, 7 of the 10 top snippets carried a 2026 date. Same engine, two very different appetites for new!

How does Google know the appetite? Both Glue and Intent Brain carry `FreshnessCounters` that sum result satisfaction by the age of the date shown in the snippet: under a day, 3, 30, 90 and 365 days, plus a counter over all dated results. That date is `bylineDate`. Compare the buckets and you get the query's need for freshness, which RESONEO reads as a broad intent called FRESH. The uncomfortable corollary from our notes: a page with no visible date probably just drops out of these counters.

Freshness turns up in two more places:

- **Repeat queries.** Repeat a search the next day and already-viewed results move down (`ViewDemotionTwiddlerParams`) while the freshest move up (`FreshestTwiddler…`).
- **The meaning of the query.** KScorer runs query interpretations through twiddlers; a freshness boost lifts the meaning that's in the news, a periodicity boost the one that returns on a date, like Black Friday. On launch day of a new Jaguar, the car can beat the animal. RESONEO: "A twiddler can therefore flip the meaning Google settles on, and the whole SERP with it."

## Which dates Google keeps for a page

So which date is "the" date? None of them! The 2024 leak describes four:

| Date (2024 leak) | What it is | Who sets it |
|---|---|---|
| `bylineDate` | "the date that will be shown in the snippets in web search results" | the site |
| `syntacticDate` | a date written in the URL or title; "never later than the crawl time" | the URL and title |
| `semanticDate` | estimated from the content, anchors and related documents | Google |
| `lastSignificantUpdate` | a combined date for the "Date Last Modified" search filter | Google |

Andrey Dzhilavdarov's Russian-language course on Google SEO, which I lean on throughout, says it in one line: you sign the byline, the system computes the content date, and it may disagree with you. Next to `lastSignificantUpdate` the leak also keeps a "pure shingle-based signal", a date derived from which pieces of text actually changed.

Patents add mechanics (a patent shows a mechanism was described, not that it's live). One splits a page into fragments: a new fragment gets today's date, a fragment seen before keeps its original one. So republished old text doesn't become new. Another weighs a page's accumulated statistics by how much each version differs from the current one: a small edit keeps almost everything, a full rewrite starts from zero. Hold on to that one.

And "first indexation"? A historical Google patent dates a document's appearance four ways: first indexing, first link, domain registration, first news mention. In the leak, `hostAge` is the earliest first-discovery date among a host's pages, counted from discovery rather than registration, and it feeds the sandbox for fresh *spam*, not for every new site.

One contradiction worth knowing: the course treats freshness predicted from URL patterns as a live signal, but in the leak that field, `freshnessEncodedSignals`, is marked "This field is deprecated".

## Why "just bump the date" backfires

Now the question I started with. The leak has a module literally named `QualityTimebasedDateUnreliability`: "Unreliable dates signals per URL". Its default is reassuring: "If scores are missing, the date is reliable by default." But the scores are computed "per Domain+Petacat", per domain within a topic category of Google's taxonomy (big sites span up to 1,500 of them), and each URL takes the maximum. The ingredients include `dateVsContentageDistributionSkew`, how far dates diverge from the content's age. Next door sits a page flag, `isPageWithFreshRepeatedDates`: "fresh repeated date sequence".

Read that with a bulk date update in mind. A fresh date on old content *is* a skew between date and content age. Our interpretation, not Google's wording: the habit likely becomes a property of the site within a vertical and undermines the dates of every page in that category. Not one page. The whole section!

The course's myth registry rates 92 of 129 popular claims about Google as refuted, and number 75 is exactly this one: "update the date and the page becomes fresh". Refuted, because a fresh date on old content is a discrepancy assessed as date unreliability. The crawl side agrees, in the course author's reading: a date change without a content change isn't an "interesting" change, and a fake `lastmod` devalues the signal while an honest one saves crawling.

> **Is it a penalty? Honestly, unclear** — Some sources call artificial refreshing a "penalty" (a Telegram channel citing David Baum, for one). The leak shows something narrower: a score marking your dates as unreliable, with no penalty field in the descriptions. Then again, a missing field doesn't disprove a penalty. Both versions agree on one thing: the date stops working for you.

The opposite extreme isn't safe either. Remember the version-difference patent: rewrite a working page from scratch and its accumulated statistics start over. The course puts "a fresh date without changes" and "a radical rewrite of a working page" side by side on its list of five mechanism-level mistakes. The honest path is narrow: change what actually aged, and date that.

## The freshness cliff in AI answers

Here freshness gets loud. On 29 May 2026 the DrMax SEO channel, citing David Baum, wrote about a documented three-month cliff: content not updated recently drops out of AI answers fast. The numbers, as retold there:

- **Salespeak:** 50% of material cited by AI is younger than 13 weeks.
- **35 million AI Overviews analysed:** AI-cited content is on average 25.7% fresher than content in traditional organic results.
- **David Baum:** a 60-day refresh cycle on the top 20 pieces got 7 of them cited after the first update; none had been cited before.
- **Trakkr, "The Half-Life of AI Citations"** (857k reports, 108k citations, 8 models, 10 months): median citation life **0 days**, mean 6.8 days, 73.5% of citations appear once and vanish.

All of it is third-party or vendor data on different samples. "90 days" is not a threshold Google announced, and RESONEO's schema data neither confirms nor refutes it. My read: the direction is believable, the number is soft. (The Trakkr figure is the scary one! A citation behaves less like a trophy and more like a stream you keep re-earning. I covered measuring that in [the post on AI-citation tracking](/blog/ai-citation-tracking/).)

Google's own plumbing is less dramatic but more concrete. Per RESONEO's October 2026 piece on AI Overviews and AI Mode, the system writing those answers attaches `FreshnessSignals` to every result, and the answer-defect grid (200+ types by RESONEO's count) includes `STALE_SNIPPET`. In shared Gemini conversations RESONEO studied, 37% of Gemini's searches added a year that wasn't in the question (13–17% in August–October 2026). Hard to read that as indifference to dates!

The channel also sorts content into four states, its own model, not Google's: recently updated, correctly static evergreen, artificially refreshed, abandoned. Only the third is self-inflicted, and the second, by the channel's account, is neutral: a page that is still true doesn't need a fake update.

## An honest date: what to actually do

The checklist I ended up with. The last item is my opinion; the rest comes from the sources.

- [ ] **One visible date in the HTML**, near the title. `bylineDate` is what snippets show and what the satisfaction counters use, and AI crawlers read HTML: a date living only in the CMS database passes no signal. Check that the date and the new paragraph survive in a bot's plain-text view; our free [AI view tool](/tools/ai-view/) shows exactly that.
- [ ] **Change `dateModified` and `lastmod` only on a significant change**: a re-sourced statistic with its year, a new example, a fixed fact, an "Updated [month year]" note. Not a typo, not a swapped stock photo.
- [ ] **Update the fragments that aged** instead of rewriting a page that works.
- [ ] **Never bulk-bump dates.** The unreliability score is per domain and category; one lazy batch can cost a whole section.
- [ ] **Don't republish old text as new.** A fragment seen before keeps its original date.
- [ ] **Wait at least one recrawl before judging**, about four weeks for text per the course author; the last crawl date is in URL Inspection.
- [ ] **Keep the visible date, the markup and `lastmod` telling one story.** My rule, not a leaked field: if they disagree, one of them is lying. And markup is no road into AI answers; Google says structured data isn't required for generative AI search, its value is rich results.

For when a page needs a refresh, a rewrite or a merge, see [content decay in the AI-search era](/blog/content-decay-refresh/).

## In short

- Twiddlers re-rank already-scored results; RESONEO counted 120, from names, without weights.
- FreshnessTwiddler is in the 2024 leak; it reads page dates and a content date Google computes itself.
- Google keeps four dates per page and scores how unreliable they are, per domain and category.
- A new date on unchanged content feeds that score; a separate penalty is unproven.
- The AI "90-day cliff" is a third-party pattern: plausible direction, soft number.

So, can we just bump the dates? Sure. Google will notice, just not the way you hoped.

## Sources

- RESONEO (Olivier de Segonzac), "Twiddlers", November 2026 — [think.resoneo.com/twiddlers](https://think.resoneo.com/twiddlers/)
- RESONEO, "Google leak part 2: understanding the twiddler framework" (the 2018 Twiddler Quick Start Guide), 31 October 2024 — [resoneo.com](https://www.resoneo.com/google-leak-part-2-understanding-the-twiddler-framework/)
- RESONEO, search intent and Intent Brain (byline-age counters), October 2026 — [think.resoneo.com/search-intent](https://think.resoneo.com/search-intent/); Navboost and Glue schemas — [think.resoneo.com/navboost-glue](https://think.resoneo.com/navboost-glue/); KScorer — [think.resoneo.com/kscorer](https://think.resoneo.com/kscorer/)
- RESONEO, sources of AI Overviews and AI Mode, October 2026 — [think.resoneo.com/ai-overviews-ai-mode](https://think.resoneo.com/ai-overviews-ai-mode/); Gemini conversations, October 2026 — [think.resoneo.com/gemini-convo](https://think.resoneo.com/gemini-convo/)
- US v. Google, notes of a call with Nayak (PXR0357), 31 January 2025 — [justice.gov](https://www.justice.gov/atr/media/1398866/dl)
- Google API Content Warehouse leak, May 2024: `bylineDate`, `syntacticDate`, `semanticDate`, `semanticDateInfo`, `lastSignificantUpdate`, `datesInfo`, `QualityTimebasedDateUnreliability`, `QualityTimebasedPetacatDateUnreliability`, `QualityTimebasedPageType`, `freshnessEncodedSignals`, `hostAge` (descriptions checked 3 and 14 November 2026)
- Andrey Dzhilavdarov, course "SEO for Google from primary sources", lessons 18–20 and the myth registry, November 2026 — [dzhilavdarov-seo.ru](https://dzhilavdarov-seo.ru/app/)
- DrMax SEO channel, post of 29 May 2026 (citing David Baum, Salespeak and an analysis of 35 million AI Overviews) — [t.me/drmaxseo](https://t.me/drmaxseo)
- Trakkr, "The Half-Life of AI Citations", 2026 — [trakkr.ai](https://trakkr.ai/trakkr-research/citation-decay/answers/how-long-do-ai-citations-last)

## FAQ

### What is a Google twiddler?

A small program in Google's Superroot that takes an already ranked list of results inside one corpus (web, images or video) and recommends boosting, demoting, filtering, capping or spacing out results; a framework reconciles the recommendations. An internal 2018 Google document counted more than 65 active in web search alone; RESONEO counted 120 in November 2026, working from schema type names only.

### Does Google have a freshness twiddler?

Yes. FreshnessTwiddler is named in the 2024 API leak: the field datesInfo is 'Used in FreshnessTwiddler' and semanticDateInfo holds data 'required by the freshness twiddlers'. RESONEO's 2026 catalogue adds a second version and a separate date score with per-query parameters. Weights and thresholds are not public.

### Which dates does Google store for a page?

The 2024 leak describes four: bylineDate (the date shown in snippets), syntacticDate (a date in the URL or title), semanticDate (estimated from the content, anchors and related documents) and lastSignificantUpdate (used for the Date Last Modified search filter). Next to them sits a per-URL score of how unreliable those dates are.

### Will changing the date make my page fresher?

Not by itself. In the leak, a mismatch between dates and content age feeds an unreliable-dates score that is computed per domain and topic category, so the habit can undermine the dates of a whole section. Change the visible date, dateModified and lastmod only when the content changed significantly.

### Is there really a 90-day freshness cliff in AI answers?

It is a pattern reported by third parties (Salespeak, an analysis of 35 million AI Overviews, David Baum, Trakkr's citation half-life study), not a threshold published by Google. RESONEO's schema data neither confirms nor refutes the 90 days. The trend looks plausible; the exact number is soft.

