BERT doesn't read your keywords. BERT reads the relationships between your words. That distinction is the single most important concept in modern on-page optimization, and almost every local SEO guide skips it.
Three vectors determine how BERT scores your content for a given intent: the Contextual Vector (meaning), the Positional Vector (proximity between service and geography), and the Segment Vector (structural patterns like Q&A). Each is an engineerable lever with clear before/after rewrites.
This is the BERT optimization deep dive for local SEO, part of our Google Maps SEO in 2026 cluster.
Run the free GeoGrid scan to see how your current BERT-scored content is performing across your service area.
What BERT Actually Does
BERT is Google's bidirectional language model, deployed since 2019 and continuously upgraded. Unlike older keyword-matching models, BERT reads text the way humans do: given a sequence of words, it predicts what words would reasonably appear around them.
The Mental Model
Older ranking logic asked: "Does this page contain the phrase 'roofer Las Vegas'? How many times?" BERT asks: "Given the surrounding context, what's the probability that this page is genuinely about roofing services in Las Vegas?"
The probability comes from relationships, what verbs co-occur with "roof"? Is the word "Las Vegas" used in a location-context or a customer-origin-context? Does the sentence structure resemble a service description, a news article, or a product listing?
All of those are engineerable signals.
Why This Matters for Local
Two pages can contain identical keywords and rank very differently based on BERT scoring. A 400-word page stuffed with "roofer near me Las Vegas roofing contractor" ranks worse than a 400-word page that uses the same keywords exactly once but surrounds them with contextually appropriate verb-noun pairs.
Vector 1: The Contextual Vector (Meaning Score)
The Contextual Vector measures whether your target keyword appears in a context that matches Google's expected meaning for that keyword.
The Principle
Surround every service keyword with high-probability verb-noun pairs that signal the correct interpretation of the keyword.
Before / After Examples
Example 1: Water Heater Repair
- Low Contextual Vector: "We do water heater repairs."
- High Contextual Vector: "Our licensed technician diagnosed the faulty heating element, drained and flushed the tank, replaced the corroded anode rod, and re-pressurized the water heater in under two hours."
Both sentences contain "water heater." The second surrounds it with "technician," "diagnosed," "heating element," "drained," "flushed," "replaced," "corroded," "anode rod," "re-pressurized." All of those are high-probability co-occurrences for water heater service, not water heater product listings. BERT scores the second sentence much higher for service-intent queries.
Example 2: Roof Replacement
- Low: "We offer roof replacement services."
- High: "Our crew installs Owens Corning asphalt shingle roof replacements, strips the existing tear-off, deploys synthetic underlayment, re-flashes the vents, and completes most residential projects in one to three days."
The high version surrounds "roof replacement" with specific product brands, specific installation verbs, specific materials, and specific timelines. BERT reads this as authentic service description, not a thin landing page.
Example 3: Personal Injury Law
- Low: "We handle personal injury cases."
- High: "Our firm negotiates settlements, litigates trials when necessary, files pleadings, takes depositions, and has recovered over $X million in personal injury case verdicts and settlements for clients across Nevada."
Same keyword. Very different BERT score.
The Pattern
Every service keyword should appear in a sentence that also contains:
- A specific verb matching the service ("installs," "diagnoses," "litigates," "replaces")
- A specific noun matching the service ("heating element," "flashing," "settlement")
- At least one concrete modifier (brand names, product categories, specific timelines)
This is the BERT contextual vector in practice.
Vector 2: The Positional Vector (Proximity Score)
The Positional Vector measures the token distance between your service keyword and your geographic keyword. Tight proximity = strong signal. Loose proximity = weak signal.
The Principle
Keep service keyword and geographic keyword within 2–5 tokens of each other. Minimize the gap wherever possible.
Before / After Examples
Example 1: Roofing in Las Vegas
- Loose Positional Vector (20+ tokens apart): "We offer comprehensive roofing solutions. Our family-owned company has served the community since 1998. We love serving the Las Vegas area and providing quality work to our neighbors."
- Tight Positional Vector (2 tokens apart): "Expert roofing services in Las Vegas, serving Henderson, Summerlin, and Paradise homes."
In the loose version, "roofing" appears in sentence 1 and "Las Vegas" in sentence 3. BERT's proximity calculation degrades rapidly across sentence boundaries. In the tight version, "roofing" and "Las Vegas" are 2 tokens apart. Much stronger signal.
Example 2: Emergency Plumbing in Henderson
- Loose: "Our company provides emergency plumbing services around the clock. Whether it's a burst pipe or a flooded basement, we respond quickly. We've been proudly serving businesses and families in Henderson and the greater Las Vegas valley for over a decade."
- Tight: "24/7 emergency plumbing in Henderson, Las Vegas, and Boulder City."
The tight version sacrifices some narrative warmth but wins on BERT scoring. Use the tight pattern in H1, meta title, meta description, and the first sentence of each section. Use the loose pattern sparingly in supporting paragraphs.
The Compound Effect
On a 2,000-word service page, every sentence where service and geography sit 2 tokens apart pushes the positional vector up. Every sentence where they sit 20 tokens apart drags it down. Dozens of such sentences compound into the page's overall positional score.
Where to Apply Tight Positioning
Priority order:
- H1: "Roofing Services in Las Vegas" (not "Roofing Services, Our Las Vegas Home")
-
-
- First paragraph, first sentence
-
- H2 and H3 headings: "Emergency Roof Repair in Summerlin"
- FAQ question headings: "Do you offer emergency roofing in Henderson?"
- Alt text on images: "Las Vegas roofing crew installing asphalt shingle roof"
Support body copy can relax the positioning for narrative flow, but the structural elements above should always use tight positioning.
Vector 3: The Segment Vector (Structure Score)
The Segment Vector measures whether your content uses structural patterns that BERT recognizes as authoritative. Q&A blocks, tables, definition lists, and clearly-delimited sections all score well.
The Principle
BERT is trained on structured content. Q&A especially primes the model to accept the answer as a factual definition of the topic.
Why Q&A Wins
Consider these two formats:
Paragraph format:
"Emergency plumbing is a service we offer for situations like burst pipes, water heater failures, or sewer backups. We respond within one hour across Las Vegas."
Q&A format:
"Do you provide emergency plumbing in Las Vegas?
Yes, we offer 24/7 emergency plumbing across Las Vegas, Henderson, Summerlin, and Paradise. Response within one hour for burst pipes, water heater failures, and sewer backups."
Identical information. BERT reads the Q&A version as a more authoritative answer because the question segment signals that a factual definition follows. Wrap the Q&A in FAQ schema JSON-LD and Google extracts it for "People also ask" rich results and AI Overview citations.
FAQ Schema as a BERT Amplifier
Every FAQ you add to your page benefits three times:
- BERT scores the Q&A segment higher than the equivalent paragraph
- Google may extract the answer into search result rich features
- AI Overviews pull from FAQ-schema content when citing authoritative sources
5–8 schema-wrapped FAQs per service page is the working target.
The Structural Patterns That Score Well
Ranked by BERT Confidence:
| Priority | Structural Element | Why |
|---|---|---|
| 1 | Tables | Extreme confidence. Cell content is treated as discrete data points. |
| 2 | Unordered and ordered lists | Very high. Each list item is a discrete statement. |
| 3 | H2 and H3 headings | High. Signals topic boundaries and semantic structure. |
| 4 | Definition lists (<dl>) |
High. Rarely used, therefore disproportionately rewarded. |
| 5 | Q&A with schema | High. Explicit question-answer pattern. |
| 6 | Paragraph prose | Standard. Passes, but doesn't amplify. |
The High Confidence Zone Hierarchy
Where you place your keywords matters more than how many times you use them. Priority placement:
Zone 1: Tables (Extreme Confidence)
If you need to express pricing, timeline, comparison, or any structured data, use a table. Every cell BERT reads as a discrete data point with high semantic weight.
Example pattern: a pricing or service-comparison table with your service keywords in the left column and details in the right column.
Zone 2: Lists (Very High Confidence)
Bulleted or numbered lists convert prose into discrete statements. Each bullet is a separate BERT-scored segment.
Example: "Our Las Vegas roofing services include:
- Asphalt shingle installation across Henderson and Summerlin
- Tile roof repair for flat-roof commercial properties in Paradise
- Storm damage restoration throughout the greater Las Vegas valley"
Three bullets, three strong BERT signals, each with tight positional vector between service and geography.
Zone 3: Headings (High Confidence)
H1, H2, H3. Each heading signals a topical boundary. Include service + geography in at least 30% of your H2 and H3 headings.
Zone 4: Definition Lists (High Confidence, Rarely Used)
The <dl> element is underused on the web. A definition list with service terms and explanations scores well and also creates interesting visual breaks.
Zone 5: FAQ Q&A with Schema (High Confidence + Rich Result Eligibility)
Covered above. The BERT amplifier plus rich result eligibility makes this the highest-ROI structural element to add to any service page.
The Information Gain Layer
BERT scoring compounds with the Information Gain Patent (#11,366,956). Google detects whether your content adds new information versus rearranging existing information. Reworded boilerplate gets discounted regardless of BERT scoring.
Why This Matters for Neighborhood Spokes
20 neighborhood spoke pages that paraphrase the same 800 words get Information Gain penalties across the board. Even if each page has perfect BERT optimization, the underlying content is low-information-gain and Google deprioritizes all of them.
The 80% unique content rule for neighborhood spokes is not aspirational, it's the Information Gain threshold. Real landmarks, real microclimate, real cost data, real neighborhood-specific FAQs. See the Hub-and-Spoke Silo guide for the spoke page requirements.
The Combined Rule
Every service and spoke page needs:
- BERT-optimized structure (three vectors working)
- Information Gain compliance (80%+ unique content)
Missing either is a cap on your ranking ceiling.
BERT Optimization Worked Example: A Full Service Page
Concrete example of what a BERT-optimized service page looks like in practice, section by section.
H1
Low: "Roofing Services" (no positional vector)
High: "Las Vegas Roofing Services: Licensed, Insured, 24/7 Emergency" (tight positional vector, specific modifiers)
First Paragraph
Low: "Welcome to ABC Roofing. We're a family-owned business that has been serving the Las Vegas community for years, and we're passionate about delivering quality workmanship to every customer we work with."
High: "ABC Roofing installs, repairs, and replaces asphalt shingle, metal, and tile roofs across Las Vegas, Henderson, Summerlin, Paradise, and Centennial Hills. Licensed and insured in Nevada since 2008, with a 20-year workmanship warranty on every new roof."
H2 Section
Low: "Our Services"
High: "Roofing Services We Provide in Las Vegas"
List Under H2
- Asphalt shingle roof replacement across Las Vegas
- Metal roof installation for commercial properties in Henderson and Summerlin
- Tile roof repair for Paradise and Spring Valley homes
- Emergency leak and storm damage restoration throughout the Las Vegas valley
Four bullets, each tight positional vector, each specific service + neighborhood pairing.
Schema-Wrapped FAQ
Do you provide emergency roof repair in Las Vegas?
Yes, we offer 24/7 emergency roof repair across Las Vegas, Henderson, Summerlin, and Paradise. Response within two hours for active leaks and storm damage. Licensed and insured in Nevada.
Wrapped in FAQ schema JSON-LD.
Comparison Table
| Service | Starting Price Range | Typical Timeline | Service Areas |
|---|---|---|---|
| Asphalt shingle replacement | $8K–$18K | 1–3 days | Las Vegas metro |
| Metal roof installation | $14K–$32K | 2–5 days | Las Vegas, Henderson |
| Tile roof repair | $400–$2,500 | Same day | Paradise, Spring Valley |
The table is a high-confidence BERT zone that also delivers practical value to the reader.
BERT Optimization FAQ
What is BERT in SEO?
BERT (Bidirectional Encoder Representations from Transformers) is Google's bidirectional language model deployed since 2019. It reads the relationships between words rather than just matching keywords. For local SEO, BERT scores your content on contextual, positional, and segment vectors, all of which are engineerable.
How does BERT affect my Google Maps ranking?
BERT-scored on-page content is a stage-2 Map Pack selection signal. Within the K-cluster candidate pool, stronger BERT scoring increases your probability of landing in the top 3. The effect compounds with S2 Occupancy, Entity Trust, and review signals. See the ranking factors article for factor weights.
What's the single biggest BERT mistake in local SEO?
Loose positional vectors, having service keywords and geographic keywords 15+ tokens apart across paragraphs. The fix is rewriting H1s, meta titles, first paragraphs, and H2/H3 headings to put service and geography within 2–5 tokens of each other.
Can I just stuff keywords to boost BERT scores?
No. BERT detects stuffing as unnatural co-occurrence patterns and penalizes rather than rewards it. The right approach is natural, varied phrasing with strategic tightness in structural elements (H1, meta, headings) and looser narrative in body copy.
Do I need to rewrite every page on my site for BERT?
No, but every page that matters commercially should be audited. Start with your service pages (non-geo), then your city hub, then neighborhood spokes. Home page and about page are lower priority.
How does BERT interact with AI Overviews?
BERT is one of the models Google uses to classify content for AI Overview citation. Well-BERT-structured content (Q&A blocks, tables, lists, tight positional vectors) is more likely to be cited in AI Overviews than flat prose on the same topic.
What tools can measure BERT scoring?
No public tool directly measures BERT scoring. Proxy measurements: ranking movement after rewrites, Google Search Console query CTR changes, GeoGrid coverage shifts after rewriting service pages. Run a GeoGrid scan before and after a BERT rewrite to see the effect.
Is FAQ schema a BERT thing or a separate feature?
Both. FAQ schema triggers rich result eligibility (separate Google feature), and Q&A structure strengthens BERT segment vector scoring (underlying model improvement). Adding FAQ schema to a page with Q&A content helps on both surfaces.
Next Step: Audit Your Current BERT Scoring
You can't directly measure your BERT score, but you can measure the ranking outcome. Run a free GeoGrid scan before any BERT rewrite to establish baseline. Re-scan after the rewrite to see the movement.
Thirty seconds to start. Heatmap by email in two minutes. Grid coverage, competitor positions, dollar leak estimate.
If your scan reveals weak coverage across your service area and your vertical plus territory is open, the Maps Domination Programâ„¢ includes BERT-optimized service page rewrites as part of the 12-week protocol. Top-3 in 12 weeks or you don't pay the success fee.
Methodology from The Google Maps Domination Playbook by Nick Thompson. For the broader framework, see Google Maps SEO in 2026, Google Maps Ranking Factors 2026, and S2 Cells and Google Maps.
External references:
