Discussion
  • Enterprise SEO
  • Search Automation

Titles and meta descriptions across millions of pages: where has rules-based templating beaten LLM generation, and where has it lost?

Published 1 reply

Discussion starter: an enterprise SEO question, posted with an answer from Niraj Raut to open the thread. If you have dealt with this on a site, reply with what you saw, especially where it differs.

Disclosure: Niraj Raut, who posted this answer, runs the SEO consultancy linked at the end of it.

1 reply

  • admin
    Admin
    Original poster

    Rules-based templates have the edge where a page’s value is carried by structured facts (price, availability, counts, dates, location) and where the title should mirror a stable query pattern. LLM generation has the edge where the distinguishing information sits in unstructured text, where the source data is messy, and in languages where joining strings together breaks the grammar. At millions of pages, the design I would trust is a hybrid: rules own the facts and the query pattern, and a model fills a small descriptive slot or cleans the inputs, behind automated checks with a template fallback.

    Two qualifications. Public controlled comparisons of LLM-written and templated titles are thin, so what follows is mechanism and test design rather than a published league table. And titles and meta descriptions are different problems: Google says snippets are primarily created from page content and that it sometimes uses the meta description instead, which changes what a generation budget should buy.

    What Google does with generated titles and descriptions

    The <title> element is an input, not the output. Google’s title link documentation lists the sources it draws on for the title shown in results. They include the title element, the main visible title, headings such as the H1, og:title, other large and prominent text, other text on the page, anchor text on the page, and text in links pointing to the page. Among the common issues it lists as reasons for a different title link is “micro-boilerplate”: repeated boilerplate text in the titles of a subset of pages. Its example is a TV site whose season pages shared one title, where Google’s title links can add the season number to tell them apart.

    That is the core risk for templates at scale. A template whose output is identical or nearly identical across sibling pages produces micro-boilerplate by design and invites Google to write its own title. Because Google also draws on the visible heading and other on-page text, a generated title that says something different from the H1 looks more exposed to a rewrite. That is an inference from the documented sources rather than a documented rule, but it argues for generating the title and the H1 from the same data, whichever method you use.

    Descriptions work differently. Google’s snippet documentation says snippets are primarily created from page content, and that Google sometimes uses the meta description when it might describe the page more accurately than on-page text. It addresses scale directly: for large database-driven sites such as product aggregators, hand-written descriptions can be impossible, and “programmatic generation of the descriptions can be appropriate and are encouraged”. It does not prefer templates or models. It asks for descriptions of the specific page, and says identical or similar descriptions on every page are not helpful when individual pages appear in results.

    Separate the input from the output:

    You control the <title> element and the meta description. Google controls the title link and the snippet. Search Console reports neither the displayed title nor whether it was rewritten, so any comparison of generation methods needs a SERP sample that records what searchers actually saw.

    Where rules-based templates win

    • Facts that change. Prices, stock, listing counts, dates and ratings live in the database and change daily. A template renders them from live data. A model-written title is a batch artefact: “from $149” or “only 3 left” goes stale when the data changes unless you regenerate on every change, and at millions of pages that is an expensive pipeline to keep correct.
    • Stable query patterns. On category, location and attribute pages, demand tends to follow fixed patterns such as “[property type] for sale in [suburb]”. A template can mirror the pattern Search Console shows for that page type. Models tend to paraphrase, and a creative synonym repeated across 400,000 pages is a sitewide relevance change nobody decided to make.
    • Clean testing. Changing one template variable across a random half of a page type is a clean split test. Model output varies wording, order and length page by page, so a win or a loss is hard to trace to anything you can reuse.
    • Accuracy and compliance. A template cannot invent a feature, a discount or a “best” claim. In regulated categories, or anywhere a false claim creates consumer-law exposure, auditable output is worth paying for.
    • Cost and change control. Templates cost almost nothing per page to run. Model output carries a cost on every generation, plus storage, versioning and QA. A model upgrade or prompt change is a sitewide title change, so it needs the same release discipline as a template change.

    Where templates lose

    • Siblings the fields cannot tell apart. When thousands of pages share the same structured values (same make, model and city; same product type and brand), the template produces near-identical titles. What distinguishes them, such as condition, a notable feature or a specific use, sits in free text the template cannot read.
    • User-generated or seller-written sources. Forum threads titled “Help!!”, listings typed in capitals, product names stuffed with codes. Rules can clean the patterns you anticipate. A model copes better with the long tail of mess, provided its output is checked.
    • Languages with rich grammar. String concatenation breaks on grammatical case, gender and plural forms. “{count} {category} in {city}” works in English and fails in languages where the place name or the noun changes form. Localisation libraries handle plurals and some gender rules, but case endings on place names are harder. This is a strong case for model help, with native-speaker review of a sample either way.

    Hypothetical example:

    A property portal has 60,000 suburb pages and 3 million listing pages. On suburb pages, a template such as “Houses for sale in {Suburb} {State} {Postcode}: {count} listings” mirrors the demand pattern, keeps the count live and leaves a model nothing useful to add. On listing pages, structured fields produce “{beds} bed {type}, {street}, {suburb}”, which is accurate but identical for dozens of units in one building. A short model-written descriptor taken from the agent’s text (“north-facing, walk to station”) separates them, but only after checks strip any claim, price or feature not present in the source data. The split is by template: rules alone on suburb pages, rules plus a checked descriptor slot on listings.

    A hybrid design that holds up at millions of pages

    The design I would build treats the model as a writer of components and candidate templates, not as a per-page author:

    Structured fields build the title skeleton, with facts in fixed slots rendered from live data
    ↓
    Are sibling pages indistinguishable from the fields alone?
    ↓
    If yes, a model writes a short descriptor once per entity from its source text. The descriptor is stored, versioned and regenerated only when that text changes.
    ↓
    Automated checks: pixel length, required tokens present, no number or claim absent from the source, correct language, not a duplicate of a sibling, no banned terms
    ↓
    Pass: publish with the descriptor. Fail: fall back to the template output.
    ↓
    Monitor displayed titles, duplicate rate and CTR by template and cohort

    Three controls keep this manageable. First, a golden set: a few hundred pages per template with human-approved titles, against which every prompt or model change is scored before release. Second, pinned model versions, because an unpinned model update changes the output of every regeneration. Third, staged releases: new output goes to a small random slice of a template first, which also gives you a ready-made test group.

    Models are also useful one level up, proposing candidate templates for each page type that people review and then run deterministically. That keeps the testability and change control of templates while using the model for what it is good at, which is suggesting wording. Query data sharpens the rules further: where Search Console shows how people search for a page type, rules can choose which attribute leads the title. The long tail often has little or no query data, and that is where a model working from page content has more to offer.

    Keep the pipeline inside Google’s policies. Google’s guidance on AI-generated content treats appropriate use of automation as acceptable and draws the line at content generated primarily to manipulate rankings. Titles and descriptions on their own are an unlikely place to cross that line. The exposure sits in pipelines that also write body copy for millions of pages, which is where the scaled content abuse policy comes into play.

    Signals that tell you which way a template should go

    Signal in your data Points towards Why
    Much of the title’s useful content comes from volatile fields (price, stock, counts) Template Model output goes stale between regenerations
    Search Console queries for the page type follow one or two patterns Template Mirroring the pattern beats paraphrasing it
    High near-duplicate rate among sibling titles Template plus a checked descriptor slot The fields alone cannot separate the pages
    Displayed titles often differ from your title element Diagnose first: note which source Google used If Google is using your H1, fix consistency before changing generators
    Source text is user- or seller-written Model with checks Rules cannot anticipate the long tail of mess
    The template runs in languages with case or gender inflection Model, or proper localisation rules, with native review Concatenation breaks grammar
    Snippets rarely match your meta description Cheap template plus a clear on-page summary Generation spend will not reach searchers

    How to test it on your own templates

    1. Measure the current state per template. Sample a few hundred URLs from each template, stratified by traffic, and record the displayed title and snippet for their main queries. Work out the rewrite rate and the near-duplicate rate (titles identical to a sibling’s once the location or ID token is removed). Templates with high rates are where the rules have run out of distinguishing information.
    2. Pick one template and randomise. Split its pages into control and variant groups matched on traffic and category. Change only the title, or only the description, on the variant.
    3. Start the clock after recrawl, not after deployment. A new title cannot affect anything until Google has refetched the page, and on a site with millions of URLs, long-tail pages can take weeks to be refetched. Check server logs for the share of variant URLs fetched since the change, begin the measurement window when most have been, then cover several full weekly cycles outside confirmed update windows.
    4. Judge on clicks against the control. Primary metric: organic clicks for the variant against the control group. Secondary: impressions as a sign of ranking breadth, CTR at comparable positions, conversions, and the displayed-title rewrite rate.
    5. Decide per template. If the model variant wins only where siblings were indistinguishable, adopt it there and nowhere else. If results are flat, keep the cheaper and more controllable option, which is the template. If the variant loses on head templates, a likely cause is drift from the query pattern, so check the displayed titles before blaming the model.

    Check meta descriptions before spending on them

    Before generating millions of descriptions with a model, measure how often Google shows the ones you have. In the same SERP sample, compare each displayed snippet with the page’s meta description. The reasoning is that on text-light templates such as listing grids and category pages, Google has less page content to draw from, so the description has a better chance of being used. Your sample will confirm or refute that for your site. Where descriptions are used, a careful template or a checked model output is worth testing. Where they are mostly replaced, the better investment is a clear summary near the top of the page, which serves readers and gives Google better material for the snippet.

    The decision rule

    I would not choose a generation method sitewide. Score each template on four things: how much of the page’s value sits in structured fields, how often those fields change, how often Google currently rewrites or duplicates its titles, and how many languages it runs in. Templates keep the pages where facts and query patterns carry the title. A model earns a slot only on templates where the rewrite and duplicate rates show the rules have run out of distinguishing information, and only after a split test on that template beats the rules.

    Need help with this on your own site?

    Niraj Raut works with businesses in Nepal, Australia, the UK, Europe and the US on technical SEO, ecommerce SEO, local SEO and AI search.

    See SEO services

Add a reply

Replies are open to approved members. Log in or apply to join.