Google doesn't read your page with its smartest model. It can't afford to.

Executive summary

A common theory in SEO circles goes like this: Google still reads web pages with "Bard", the older model, because running Gemini over the whole web would cost too much, so rankings lag behind what Google's newest AI can actually understand.

The detail is wrong. The instinct behind it is mostly right.

Bard was never a ranking system. It was a chatbot, renamed Gemini in February 2024, and it had nothing to do with how pages are scored. Organic ranking runs on its own stack: RankBrain, neural matching, BERT-based models and click-trained systems, described in Google's own documentation and in sworn testimony from the 2023 antitrust trial. But the economic argument holds up. Google's ranking systems guide says MUM, the model Google called "1,000 times more powerful than BERT" in 2021, "is not currently used for general ranking in Search." The guide doesn't mention Gemini or large language models at all. The language understanding that decides your organic position is a generation or more behind the model you talk to in the Gemini app, and the main reason is the cost of running it billions of times a day.

What Bard actually was

Google launched Bard in March 2023 as a conversational assistant. In February 2024 it was renamed Gemini, with a paid Gemini Advanced tier and a dedicated Android app. Google's stated reason was to give people "direct access to Google AI".

Nothing in that history connects Bard to page analysis. A chatbot answers one user's prompt at a time. Ranking has to score many candidate pages for every one of the billions of searches made each day, inside a response time measured in milliseconds. Those are different engineering problems, and Google has always solved them with different models. So "Google still uses Bard to read pages" is not a defensible claim. Anyone who repeats it in a pitch or a client report can be corrected with one search.

The stronger version of the argument doesn't need Bard at all.

What Google says reads your page

Google publishes a list of the ranking systems it runs. Four of them are language models or close relatives:

The guide was last updated in December 2025. That's almost two years after Gemini launched, and it still names MUM as the most advanced system while saying it stays out of general ranking.

The antitrust trial added more detail. Pandu Nayak, Google's vice president of Search, testified in October 2023 about systems including Navboost, RankBrain, DeepRank and RankEmbedBERT. Coverage of that testimony describes DeepRank as BERT applied to ranking, and RankEmbedBERT as a BERT-based model trained on search logs that re-ranks results already retrieved by older systems. AJ Kohn, who read the transcript closely, notes that RankBrain "is run on a smaller set of documents at the tail end of ranking". Some trial coverage repeats a figure of the top 20 to 30 results. We could not confirm that number against the transcript, so treat it as reported rather than established.

Taken together: the language understanding behind organic ranking is built on BERT, a 2018 research model, plus systems trained heavily on click behaviour. Google's heavy models work late in the pipeline, on a shortlist. They don't read every page.

Why: the cost of reading the web

The economics are spelled out in Google's own announcements. They are not speculation.

When Google brought BERT into Search in October 2019, it applied it to about 1 in 10 English searches in the US. The same post said some BERT models were "so complex that they push the limits of what we can do using traditional hardware," which is why Google served them on its latest Cloud TPUs. That was BERT, a model tiny by current standards, and at launch Google rationed it to a tenth of US English queries.

Five years later the same pressure showed up on the generative side. On Alphabet's Q3 2024 earnings call, Sundar Pichai said Google had "reduced costs by more than 90%" per query for AI Overviews "through hardware, engineering, and technical breakthroughs, while doubling the size of our custom Gemini model." Two details in that sentence matter. Cost per query was a headline metric for the CEO, and the Search model is a "custom Gemini model". It's a version sized for serving at Search volume, not the largest Gemini model Google offers.

Google's developer pricing shows the same split. Its API sells Pro, Flash and Flash-Lite tiers at very different prices per token, and Pro is the expensive one. Search runs at a volume no API customer comes close to. Running a frontier-size model over every candidate page for every query would cost a large multiple of what Search can carry.

So Google does what any engineering team does with an expensive component. It uses cheap systems (the index, lexical matching, link and click signals) to narrow billions of pages to a shortlist, applies heavier language models only to that shortlist, and keeps its most capable models for products where one user is waiting for one answer.

What that costs the results page

Here is the part of the original theory that holds up. If the model judging relevance is BERT-era, then relevance is judged with BERT-era understanding.

These are inferences from the documented architecture. Google doesn't publish them:

Cheap retrieval decides who gets read at all. The heavier models only see pages that already made the shortlist. A page with a precise, well-reasoned answer that doesn't use the vocabulary of the query may never reach the stage where nuance gets weighed.

Click-trained systems reward what people already click. Navboost and RankEmbedBERT both learn from user behaviour. That favours familiar brands and familiar phrasing over a better answer from an unknown site, regardless of how well the newer page reads.

Long-range reasoning is the gap. Modern LLMs can follow an argument across a whole document, notice when a page contradicts itself, or tell a first-hand account from a rewrite of one. BERT-family rankers work on much shorter stretches of text and are better at matching than at judging. Google covers part of that gap with other systems, such as helpful-content signals, spam classifiers and site-level quality signals. Those systems are blunter instruments.

The newest model shows up in a different box. AI Overviews and AI Mode do use a custom Gemini model. So the same results page can have a generative summary written by a 2024-or-later model sitting above ten blue links ranked by systems whose language understanding dates from 2019. That's why an AI Overview sometimes cites a page that ranks on page 2.

Counter-argument, taken seriously

"Google doesn't publish everything. Gemini could be in ranking without being named." True, and it's the strongest objection. Google uses large models to train and evaluate smaller ones, and a distilled model in production would not need to be called Gemini. The ranking systems guide also lists systems, not model versions. But Google's public documentation still says its most capable named model is not used for general ranking, and nothing Google has published says otherwise. Until it does, "BERT-family rankers plus click signals" is the documented position. Claims that Gemini scores pages in organic ranking are speculation.

"Older doesn't mean worse at the job." Also fair. A small model trained on billions of real queries and clicks can beat a frontier model at the narrow task of ordering ten results for one query. Engagement data captures things no language model can see. The real cost of the cheaper stack is smaller than "Google ranks with outdated AI" suggests. It shows up mainly on long, complex or unusual queries, where matching and click history matter less than reasoning.

"Cost isn't the only reason." Latency and stability matter too. A ranking model has to answer in milliseconds and behave the same way tomorrow as today. Large generative models are slower and less predictable. Cost is the reason Google talks about most, but it's one constraint among several.

What this means for how you write

If the reader deciding your organic position is a matching-focused model working with click data, write so that the cheap stages can't misread you:

1. Use the searcher's words. Retrieval and BERT-style matching reward pages that say plainly what the query asks. Put the exact question, or something close to it, in a heading, and answer it in the first sentence underneath.
2. Keep facts close together. Short-context models score passages. A claim and its evidence three paragraphs apart may never be read together. Name the entity, the number and the source in one place.
3. Don't rely on implication. "As mentioned above" and pronouns that point back several sections are easy for an LLM and hard for a matcher. Repeat the noun.
4. Earn the click. Click-trained systems weight what people choose. Titles and snippets that match intent carry more weight than polishing paragraph 9.
5. Write the whole page for the generative layer too. AI Overviews do use Gemini-class models, and they cite passages that a modern model judges useful. A page built for both readers is plainly matched for ranking and well argued for citation. See What Actually Works for AI Citations.

What everyone is missing

There are two Googles on one results page. The AI summary and the organic list are produced by different generations of language technology with different budgets. Optimising for one and assuming the other follows is how a site ends up cited in an Overview but not ranking, or the other way round.

The cost curve is moving. BERT went from rationed to universal within about a year of launch. Pichai's 90 percent figure shows how fast serving costs fall. The gap between what ranks pages and what Google can afford to run will narrow. Pages that are clear to a cheap model and convincing to an expensive one hold up through that change.

"Google uses Bard" is a credibility trap. It's easy to disprove, and it gets used to explain ranking drops that have ordinary causes. The documented version of the argument is stronger and survives scrutiny.

Practical takeaways

1. Drop the Bard theory. Bard was a chatbot renamed Gemini in February 2024 and never ranked pages.
2. Google documents that its strongest search language model, MUM, is not used for general ranking. Its public ranking documentation doesn't mention Gemini.
3. Organic ranking runs on BERT-family models, RankBrain, neural matching and click-trained systems. Heavier models work late in the pipeline on a shortlist.
4. Cost is a documented constraint. BERT launched on about 10 percent of US English queries, and Google cut AI Overviews' cost per query by more than 90 percent while using a custom Gemini model sized for Search.
5. Write for both readers. Plain query language and facts kept together for the ranker, a well-reasoned page for the generative layer. See Semantic Authority.