Google BERT Update Explained: Writing for AI Overviews

BERT stands for Bidirectional Encoder Representations from Transformers, the natural language model Google rolled out in October 2019 to understand full sentence context rather than isolated keywords. It is still part of the infrastructure behind every Google search today, including the AI Overviews now appearing above standard results. If you are writing content for Google in 2026, you are writing for BERT, whether you know it or not, and increasingly, you are writing for the generative systems built on top of it, too.

For businesses in Northern Ireland, Ireland, and the UK, that shift has particular implications. Regional language, local intent signals, and the way people in Belfast or Dublin actually phrase a search query are factors the algorithm actively interprets. Understanding how BERT works, and how it feeds into Gemini and AI Overviews, is the foundation of any content strategy that wants to be found (and cited) in 2026.

What Is the Google BERT Update?

BERT is a natural language processing model developed by Google AI to help the algorithm understand the full context of words in a search query, rather than reading them as isolated signals. Before BERT, Google’s algorithm largely read queries left to right, treating each word as a standalone clue. BERT reads in both directions at once, taking in the whole sentence before interpreting any individual word within it.

The practical result was immediate. Queries built around prepositions, negations, or implied context, the kind that change meaning entirely with one small wording shift, became far more accurately matched to relevant content. A search for medicine “for someone else” stopped returning generic pharmacy pages and started returning content about collecting prescriptions on behalf of another person. That distinction is BERT at work.

Why BERT Still Matters in the Age of Gemini

BERT is often discussed as a 2019 event that changed things and faded into the background. That framing understates it. BERT is not a feature Google can switch off. It is embedded in the core infrastructure Google uses to process language, and newer systems, including MUM and Gemini, are built on the same transformer architecture, extending what BERT established.

Gemini adds multimodal reasoning and the ability to generate direct answers rather than just rank pages, but it still depends on the natural language foundations BERT introduced. If you want to understand why Gemini interprets your content the way it does, or why an AI Overview picked one passage of your page over another, BERT is still the right place to start.

BERT and Regional Language: The UK and Ireland Difference

This is where BERT delivers specific value for businesses operating in the British Isles, and where US-centric SEO advice regularly falls short. UK and Irish English is not the same as American English in vocabulary, phrasing, or search intent. BERT’s contextual processing lets it distinguish between terms that overlap in one dialect but diverge in another.

A few examples show how this plays out in practice:

• “Solicitor” vs “lawyer”: in the UK and Ireland, these are distinct roles. A search for “solicitor Northern Ireland” carries a different intent from “lawyer Northern Ireland,” and BERT processes that distinction. A law firm using both terms correctly, in context, is better served than one that stuffs “lawyer” into a solicitor’s page because it has higher search volume.
• “Skip hire” vs “hire a skip”: BERT identifies both phrasings as the same intent, which means natural, conversational copy performs as well as exact-match repetition.
• “GP” vs “doctor” vs “physician”: BERT understands the UK health-search hierarchy and can match “GP Belfast” to content using “general practitioner” and “NHS” as supporting context, even without the exact keyword in every paragraph.

For SMEs across Northern Ireland, Ireland, and the UK, content written in the language actual customers use, rather than a generic keyword list, is rewarded directly by the algorithm. This is one reason ProfileTree’s SEO services start with how a business’s own customers phrase things, not with a template keyword sheet.

BERT vs RankBrain vs MUM vs Gemini: Decoding the Algorithm Stack

Google’s algorithm is not a single system. It is a stack of interconnected models, each handling a different part of how queries are processed and content is ranked.

RankBrain learns from patterns across queries. BERT understands language structure within an individual query. MUM connects knowledge across formats and languages. Gemini generates answers rather than only ranking pages. They work together, not in sequence.

“Businesses often ask us whether they should be optimising for BERT, MUM, or Gemini separately,” says Ciaran Connolly, founder of ProfileTree. “The honest answer is that you’re optimising for all of them at once, and the requirements are the same: clear, structured content that answers real questions in the natural language of your audience. The algorithm stack changes; the principle doesn’t.”

What Are Google AI Overviews, and How Do They Build on BERT?

Google AI Overviews are AI-generated summaries that appear above standard results for certain queries, drawing on multiple ranked pages and citing sources. The mechanism behind them is retrieval-augmented generation, usually shortened to RAG: Google retrieves pages it already ranks well using its normal systems, then passes those pages to a Gemini model, which drafts a summary and attributes what it used. That detail matters for anyone trying to appear in one. The overview is grounded in pages Google already ranks, so the underlying SEO still has to work first.

This is a direct extension of what BERT started. BERT taught the algorithm to understand full sentence meaning rather than keyword lists. AI Overviews take that contextual understanding and use it to draft an answer, rather than simply pointing to a page that might contain one.

Table stakes vs Featured Snippets is worth clarifying, since the two features are often confused:

According to Google’s own guidance for site owners, AI features draw on retrieval methods that surface content already sitting in Google’s search index, and Google treats optimising for generative AI results as part of ordinary SEO rather than a separate discipline. The same guidance pushes back on one specific tactic: it says content doesn’t need to be broken into small chunks purely for AI systems, since Google’s own systems are capable of reading a multi-topic page and pulling out the relevant section without that extra pre-processing. For any SME being sold an “AI chunking” package, that’s worth knowing before paying for it.

For a deeper walkthrough of how to structure a page for AI Overview inclusion specifically, see ProfileTree’s guide to appearing in Google AI Overviews.

Writing Content Google’s AI Systems Can Extract

Optimising for BERT, and by extension for AI Overviews, is not a separate task from writing good content. It is the same task, done properly.

Write for intent, not keyword density. BERT rewards content that fully addresses the meaning behind a query, not content that repeats a phrase across a fixed number of paragraphs. For “how does local SEO work for a Belfast restaurant,” useful content covers what local SEO is, what signals matter for that business type, and what a restaurant owner needs to do differently in Belfast than in a large city. Content that inserts “Belfast restaurant local SEO” a dozen times into generic advice does not serve that intent, and BERT knows it.

Treat prepositions as meaning markers. Before BERT, “SEO for manufacturers” and “SEO manufacturers” were treated as near-equivalent. After BERT, the “for” signals audience specificity. This matters in headings, introductions, and FAQs: “digital marketing for hospitality businesses in Northern Ireland” targets a different intent than “Northern Ireland digital marketing,” even with overlapping keywords.

Map entities explicitly. BERT works alongside Google’s Knowledge Graph, so content that clearly connects entities, people, businesses, places, services, is better understood than a page optimised for a keyword list. For a Northern Ireland digital agency, this means content that connects “ProfileTree,” “Belfast,” “web design,” and “SMEs” in plain declarative sentences, not as a keyword strategy but because it reflects how those things actually relate. This kind of entity clarity is also what good site structure and schema markup are built to support technically.

Answer the full query in a self-contained passage. AI Overviews and voice assistants both draw on passages that stand alone. Open each major section with a direct, two-to-three-sentence answer, then support it. This does not mean chopping content into disconnected fragments; Google itself says fragmenting content for AI systems is unnecessary. It means writing the way a good encyclopaedia entry is written: clear at the top, detailed underneath.

Match length to the question, not a target.ProfileTree’s content length guide covers this in more detail, but the short version is that word count is not the ranking factor; whether the content answers the question completely is.

The UK and Ireland Data Protection Angle

Almost every guide to writing for AI Overviews is built on US search behaviour and US data assumptions. UK and Irish businesses operate under a different framework, and it is worth understanding in broad terms rather than assuming US advice transfers directly.

UK GDPR governs how personal data is collected, processed, and used, and the Information Commissioner’s Office (ICO) is the relevant supervisory authority for most UK SMEs. This framework is primarily about personal data, not about a business’s published editorial content, so it does not give a business a general right to control whether Google’s crawlers read and summarise its public web pages. What it does affect is anything on a page that touches personal data: customer reviews with identifying details, case studies naming individuals, or contact forms that feed into AI-assisted marketing tools. If your content includes any of that, the same data minimisation and consent principles that apply to your marketing generally still apply here.

For a fuller grounding in the underlying rules, see ProfileTree’s guide to GDPR. If your business operates across the UK and EU markets, rollout of AI search features and the regulatory scrutiny around them can differ by market, so it’s worth treating this as a live area to monitor rather than a settled question, particularly if digital strategy decisions depend on it.

Should You Block AI Overviews From Your Content?

Being cited in an AI Overview is not automatically the right goal for every page. Informational content that gets fully answered inside the summary can lose the click it would previously have earned. Commercial and highly specific content, where the reader needs more than a summary to act, tends to fare better, since a citation there often brings a more qualified visitor than a broad organic click would.

Google gives site owners standard meta robots directives to control this, and they are worth understanding before assuming you have no choice:

There is no universal right answer here. The decision depends on whether a summary is likely to satisfy the reader’s need on its own, or whether it is likely to send a better-qualified visitor through. This is a judgement worth making page by page rather than site-wide, and it is the kind of decision ProfileTree talks through with clients as part of ongoing SEO work.

Measuring Success: Tracking AI Overview Visibility

The practical difficulty is that Google Search Console does not currently give site owners a separate filter for AI Overview impressions and clicks. They are merged into standard organic web results, which makes precise in-house attribution difficult without extra work.

A workable approach for most SMEs is to track a query’s overall impressions and position over time, then manually check whether an AI Overview appears for that query and whether your page is among the sources cited. It is manual and imperfect, but it is more reliable than assuming a traffic dip or rise is explained by AI Overviews without checking. Third-party rank tracking tools that specifically flag AI Overview presence can reduce the manual work at scale, though none of them offers the precision of first-party Search Console data.

BERT and Local SEO: What It Means for UK and Irish Businesses

Local SEO was one of the areas most directly affected by BERT’s improvements in contextual understanding. Before BERT, local queries worked mainly through proximity signals, NAP data, and keyword matching. BERT added a layer of intent interpretation that made local search considerably more nuanced.

A query like “solicitor near me that handles estate disputes in Northern Ireland” now triggers matching based on the legal meaning of “estate disputes,” not pages about property valuations. That precision is BERT at work, and it carries through into how AI Overviews select and cite local businesses, too.

For SMEs, the practical implication is that location pages and service pages written in natural, specific language outperform those built around keyword repetition. A page that states clearly that a business “provides SEO services to small and medium businesses across Belfast, Derry, and the wider Northern Ireland market” communicates location, service, and audience in a way both BERT and AI Overviews can process. Businesses operating across several towns or counties face a related challenge covered in ProfileTree’s guide to regional SEO for the UK and Ireland, and those competing specifically for AI-driven local visibility may find ProfileTree’s guide to AI SEO in Ireland useful background.

BERT and Voice Search: What It Means for UK and Irish Businesses

Voice search and BERT are a natural pairing. Voice queries are conversational: longer, phrased as complete questions, and dependent on prepositions and qualifiers to carry meaning. These are exactly the characteristics BERT was built to process.

Someone typing a query compresses it: “SEO agency Belfast.” Someone speaking it doesn’t: “What’s a good SEO agency for a small business in Belfast?” The spoken version carries intent signals, “good,” “for a small business,” “in Belfast,” that keyword-based systems historically struggled to weight correctly. BERT handles them directly, and the same structure that satisfies a voice assistant tends to satisfy an AI Overview.

For UK and Irish businesses, this matters because voice search reflects regional speech patterns that differ from the American English dominant in most SEO training data. A query asking about “getting the bins emptied” or “claiming back VAT” only makes sense within a regional context, and BERT’s contextual processing handles that better than any earlier version of Google’s algorithm. ProfileTree’s guide to voice search optimisation covers the technical side, including schema markup for voice-friendly content.

Building This Into Your Content Workflow

None of this requires a separate “AI content” process bolted onto existing work. It requires the same editorial discipline BERT has always rewarded, applied consistently: answer the real question, use the language your customers use, and structure the page so both a person and a machine can follow it.

Two things make that consistent rather than occasional. First, readable, well-edited copy, since AI-drafted first passes tend to default to a uniform sentence length and generic register that reads as thin to both people and ranking systems. Second, visible author and business credibility, covered in more depth in ProfileTree’s guide to E-E-A-T, since Google’s February 2026 update made author credentials a more direct ranking input.

For teams handling this in-house, this is usually a training gap rather than a tooling gap: knowing what a passage needs to do, not which AI tool to run it through. That’s the area ProfileTree’s digital training and AI training programmes are built to close for SME marketing teams.

Where This Leaves UK and Irish Businesses

BERT is not a historical event. It is part of the operating infrastructure of search in 2026, the foundation on which MUM, Gemini, and AI Overviews are built. For businesses in Northern Ireland, Ireland, and the UK, it reinforces something good content writers have always known: write clearly, write for your actual audience, and use the language your customers use. If a content strategy still relies on keyword density rather than intent coverage, it’s worth reviewing how the page reads to a person, not just to a crawler, before assuming AI Overviews are the problem.

FAQs