Why this article exists
Two things changed search in the last seven years. The first was BERT, in October 2019 — the moment Google moved from matching keywords to understanding sentences. The second was AI Overviews, the generative answer block that now sits above the blue links for a growing share of queries. Marketers I talk to know AIO exists. Almost none of them know BERT is what made AIO possible, or how the two relate to the everyday choices they make on a content brief.
That's a problem, because the rules have shifted. Old SEO rewarded keyword density and exact-match phrases. Modern SEO rewards being the page Google's models understand best on a topic — and being structured enough that a generative answer can quote you. Those are different jobs.
This piece walks the line from BERT through MUM and on to AI Overviews, then turns the theory into a working playbook. It assumes you've read the E-E-A-T piece and the Schema piece — the moves below stack on top of those.
BERT — what it is, in plain English
BERT is short for Bidirectional Encoder Representations from Transformers. Three words do the heavy lifting: bidirectional, encoder, and transformer.
A transformer is a kind of neural network introduced in a 2017 paper called "Attention Is All You Need." Its key trick is the attention mechanism — for every word in a sentence, the model can look at every other word and weigh how much each one matters for understanding the current word. Pre-transformer models read sentences left-to-right or right-to-left like a human reader. Transformers read all of it at once, and decide on the fly which other words are relevant.
BERT was Google's 2018 application of that idea to language understanding, with one important twist: it's bidirectional. When BERT encodes the word "lead" in the phrase "cost per lead in B2B SaaS 2026," it doesn't just see the words to the left ("cost per"). It sees everything in both directions before deciding what "lead" means. That's how it knows we're talking about a sales lead, not the metal, not a leash, not the verb.
And "encoder" means the model produces a numerical representation — a vector — of every word and every passage. Two passages that mean similar things end up near each other in vector space, even if they share no exact words. "Cost per lead" and "CPL by channel" land close together. "Lead" the metal and "lead" the sales prospect land far apart.
That single capability — encoding meaning instead of matching strings — is what made everything that came after possible.
From BERT to MUM to AI Overviews
BERT was the start, not the end. Google's language models have compounded since:
- BERT (October 2019) — applied first to ~10% of English queries, soon expanded to most. The first time Google could reliably handle queries with prepositions, negation, and complex intent. Search felt smarter overnight, and most marketers didn't notice.
- MUM (2021) — Multitask Unified Model — a much larger model trained on 75 languages and on text + images jointly. MUM's job wasn't to replace BERT but to handle harder queries: "I've hiked Mt. Adams; what should I do differently to prepare for Mt. Fuji next autumn?" That's a question with three intents, two locations, a temporal frame, and an implied request for advice. MUM made queries like that tractable.
- RankBrain, neural matching, and helpful-content systems have layered on top in parallel — but BERT and MUM are the language-understanding spine.
- SGE (Search Generative Experience), 2023 — Google's first public experiment with putting a generative answer above the blue links. Beta only, opt-in, and very rough.
- AI Overviews (May 2024 GA) — SGE renamed and shipped to general availability. Powered by a Gemini-family model, with the language-understanding layer below it tracing right back to BERT.
The line is straight: BERT lets Google understand the page. MUM lets it reason across modalities and languages. AI Overviews let it write you an answer using both, plus citations.
AI Overviews — what they actually are
An AI Overview is the boxed, generative answer Google now shows above (or near the top of) the SERP for a growing share of queries. It's not a featured snippet — those are quoted directly from a single page. An AIO is composed text, generated by a Gemini-family model, drawing on multiple sources that are then cited as supporting links beside the answer.
From the marketer's perspective, an AIO is two things at once:
- A click suppressor. When the user gets the answer in the box, they often don't click through. Studies in 2024–25 reported click-through compression of 30–60% on informational queries where AIO appears. That hurts.
- A citation surface. The 2–6 sources cited alongside the AIO get a tiny but real lift, plus a brand-attribution win. Being cited beats not being cited, even when the user doesn't click.
Both are happening at the same time. The strategy is to be the page that gets cited inside the AIO and ranks underneath it for the cases where the user wants more. Doing one without the other is half a strategy.
What triggers an AI Overview
AIOs don't appear on every query. Google's own pattern, observed across hundreds of queries:
- Informational and comparative queries trigger AIO most often. "What is X," "X vs Y," "how does X work," "best X for Y" — the AIO is showing up on 40–70% of these in 2026.
- Navigational queries (brand names, exact product names) almost never trigger AIO. If someone searches "anthonyligyat.com," they get the site, not an answer.
- Commercial-intent queries are mixed. "Buy [product]" tends not to trigger AIO; "best [product] for [need]" often does.
- YMYL queries are conservative. Health, legal, and financial queries get AIOs only when the answer can be drawn from very high-trust sources. Read the YMYL piece on why this is the right call.
- Localised and multilingual queries trigger AIO unevenly. EN-US is most aggressive; EN-AU and most non-English markets see AIOs less often, but the trajectory is global.
What gets cited inside an AIO answer
The cited sources beside an AIO aren't always the top blue-link results. Patterns I've watched in the wild:
- Pages that answer the specific sub-question the model is composing on. AIO is often answering a slightly narrower version of the user's query than the blue links are ranking for. Pages that include that narrower version as an explicit subheading get cited disproportionately.
- Pages with clear, scannable structure. H2s, short paragraphs, definition lists. The model is sampling passages, and passages with clear edges are easier to lift.
- Pages with credible authorship. Visible byline, an /about page that's actually about a person, sameAs to LinkedIn — the same E-E-A-T signals that affect blue-link ranking, only weighted harder for YMYL topics.
- Pages with structured data that matches the query intent. An FAQPage schema for an FAQ-shaped query, an HowTo schema for a "how do I" query, an Article schema with proper author + image. The Schema piece covers this in depth.
- Pages that cover the breadth of the topic. AIO is biased toward sources that look like they understand the whole concept, not just one slice. A 2,500-word piece often beats five 500-word pieces on the same site.
- Pages that have been around long enough to accrue trust signals. Backlinks, mentions, and dwell time all still matter. AIO didn't repeal PageRank; it added a layer above it.
The marketer's playbook
Five moves that compound. None of them are exotic. The compounding is the point.
1 · Write to entities, not keywords
Pre-BERT, the unit of optimisation was the keyword phrase. Post-BERT, it's the entity — a thing in the world that Google understands, with relationships to other things. "Cost per lead" is an entity. So is "B2B SaaS." So is "LinkedIn Campaign Manager." A page about CPL in B2B SaaS should namedrop the entities it's connected to (CAC, MQL, SQL, channel mix, attribution windows) without keyword-stuffing them. Google's models are fluent enough to read this as "this page understands the topic," not "this page is keyword-spamming."
2 · Structure for sampling
Treat every H2 as a possible AIO citation point. The model is going to look for a passage that cleanly answers a sub-question. The easier you make it to lift a clean 80–120 words, the more often you'll be lifted. Concretely:
- One idea per paragraph. Edit out throat-clearing.
- Put the answer in the first sentence after a heading. Elaborate after.
- Definition-style openers ("X is…") for explainer queries.
- Numbered steps for how-to queries.
3 · Schema as the bridge
Structured data gives the model an explicit, machine-readable answer to "what is this page about?" — saving it from inferring. Article schema with proper author, image, and dateModified for explainer pieces. FAQPage schema only for genuine FAQs. HowTo schema for procedural content. The Schema piece goes through the patterns. Match the schema type to the actual content; mismatched schema gets penalised, not rewarded.
4 · Visible authorship + sameAs everywhere
An AIO that cites a page with a clear named author and a verifiable identity is a much safer answer for Google than one citing an anonymous post. Make the byline a real link to a real /about page. Use Person schema with sameAs pointing to LinkedIn (and any other authoritative profile). The E-E-A-T piece covers this on the trust dimension; here, it's also a citation-eligibility signal.
5 · Cover the topic, not just the slice
AIO favours sources that look like they understand a subject end-to-end. Three deep articles with internal cross-links beat ten shallow articles. If you've already written a piece on CPL, the second piece should be on attribution windows and link to the first; the third should be on channel mix and link to both. You're building a topic cluster, but the goal isn't internal-linking SEO — it's giving the model evidence that this site is the place to draw from.
Mistakes I see most often
- Optimising for keyword variants Google has already collapsed. "Best CRM for small business" and "small business CRM software" are the same intent. Don't ship two posts.
- Writing to be quoted by ChatGPT instead of Google. Every AI assistant has a different sampling pattern. Optimising for one specific assistant is whiplash. The fundamentals — clear structure, credible authorship, comprehensive coverage — work across all of them.
- Adding FAQ schema to pages that aren't FAQs. The 2023–24 Google update specifically demoted this. The schema type has to match the actual content type.
- Treating AIO as something to "rank in." You don't rank in AIO. You earn citations beside it. Different mental model, different tactics.
- Panicking about CTR loss without measuring brand-attribution lift. Some queries that lose clicks gain brand-search lift two months later. Measure both before deciding the AIO killed your channel.
Practical checklist
Three categories. Pick the ones that aren't already done on your highest-traffic page, ship those first.
- ☐ Real
/aboutpage with author photo (or editorial illustration), bio, credentials, sameAs to LinkedIn. - ☐ Visible byline on every article that links to
/about. - ☐ Article schema with author, image (ImageObject with width/height/caption), datePublished, dateModified.
- ☐ H2s that are answer-shaped, not topic-shaped. "What is X?" and "How does X work?" beat "About X."
- ☐ One-idea paragraphs. The model can't lift a paragraph that's saying three things.
- ☐ Internal cross-links between topical pieces. Three deep > ten shallow.
- ☐ FAQPage schema only on real FAQs.
- ☐ HowTo schema only on real procedures.
- ☐ Image alt text that describes content, not "image of."
- ☐ A Search Console regex query filter that tracks where AIO is showing up on your terms — query patterns like
what is,how does,vs,best,for. - ☐ A monthly audit of your highest-impression queries that now show AIO. Check whether you're cited; if not, what would have to change.
Most of this list is invisible to the user and unmistakable to a model. That asymmetry is the opportunity. Spend a week on a single piece doing all of the above — including the schema, including the byline, including the topic cluster — and watch what happens to your impressions on the queries that show AIO. The compounding starts almost immediately.