Local search ranking factors

Google's local search algorithm is the system that turns a single user-typed query into a ranked, filtered, and personalised list of nearby businesses, in a fraction of a second. It is the most operationally important algorithm any local business needs to understand, and it has changed more between January 2024 and September 2026 than in the eight years before that. This is the deepest current treatment of how the algorithm actually works, the signal stack it ranks on, and what every major update from Pigeon through to the August 2026 spam update changed in practice.

The mental model: three public pillars, one silent fourth

Google's own help documentation lists three inputs the local algorithm uses to rank results: relevance, distance, and prominence.¹ That framing is real and load-bearing, and we will treat each one in depth below. What it understates, especially in 2026, is the role of a fourth class of signals Google reads continuously but does not name in the public help: behavioural data from the searches and interactions users have with listings.

What changed from 2024 to 2026

If you were last seriously across the local algorithm in 2022 or 2023, the most important shifts to internalise are these:

• AI Mode and AI Overviews retrieve from the same local index. When a query carries local intent and Google's AI surfaces respond, they pull candidates from the same Map Pack index and re-rank them with an entity overlay. The Map Pack winners are not a separate problem from the AI Overview citations; they are the same candidate set scored slightly differently.
• Multi-vector retrieval (MUVERA-style) is reshaping candidate selection. Google has not confirmed exactly where its MUVERA work is deployed, but the observable behaviour fits it: candidate selection increasingly combines lexical match (traditional keyword matching), entity match (Knowledge Graph alignment),⁴ and semantic match (vector similarity to the query embedding). A listing that wins on only one of the three is rarely a candidate against listings that win on two or three.
• Query fan-out splits user queries in AI answers. Google says AI Overviews and AI Mode may use a "query fan-out" technique, issuing multiple related searches across subtopics and data sources.⁵ A query like "best italian restaurant for date night" can become several searches (italian restaurant near me, romantic restaurant, dinner). Google has not said the Map Pack works this way, but a listing that matches more of those sub-queries has more chances to be picked up in the AI answer.
• Schema.org keeps adding types that matter for local.⁶ v30.0 (March 2026) added the Credential type and equivalence annotations to GS1, Dublin Core, and Open Graph; v29.4 (December 2025) added OnlineMarketplace and ConferenceEvent. Together they give Google more direct typing for regulated professions and specialised commerce.
• ChatGPT, Claude and Perplexity retrieve local data through their own pathways. OpenAI says ChatGPT search uses third-party search providers, and ChatGPT now draws on licensed Yelp reviews, photos and listings; Claude's web search runs on Brave Search; Perplexity runs its own search crawler. None of them documents reading Google's local index. Whether you appear in their answers is correlated with Map Pack performance but not identical.
• Photo quality appears to beat photo count. Google does not document how it assesses profile photos, but a listing with 12 high-quality recent photos tends to do better than one with 60 mixed-quality photos.
• Behavioural signal weight has grown materially. Compared with a few years ago, behavioural signals appear to carry materially more relative weight. This is the single biggest practical shift in the algorithm.

How a local query becomes a result

The end-to-end flow from typed query to ranked Map Pack runs through seven stages. The whole thing happens in a fraction of a second, but inside that window the system is doing significant work:

1. 1

   Query parsing and tokenisation

   The user query is tokenised, entities are extracted (brand names, locations, attribute words), and a normalised form is produced. "best italian rest. near me open now" becomes a structured representation: category=italian restaurant, modifier=best, time-constraint=open now, geo-anchor=user location.
2. 2

   Local intent classification

   A binary classifier decides whether this query should surface local results at all. The classifier is conservative; queries like "Italian recipes" do not trigger a Map Pack, but "Italian food" usually does. The output also includes a local-intent strength score that influences whether the Map Pack appears, how many results, and whether AI Overviews are offered alongside.
3. 3

   Geo-anchor selection

   Where to centre the search radius. On mobile this is the user's GPS coordinates. On desktop it is IP-based (less precise, typically accurate to city level). An explicit geo modifier in the query ("italian restaurant in Toronto") overrides both with the named location's centroid.
4. 4

   Multi-vector candidate retrieval

   Candidates are pulled through three parallel retrieval paths: lexical (keyword matching against listing fields), entity (matching against the Knowledge Graph for businesses of the right type within the radius), and semantic (vector similarity between the query embedding and listing embeddings). The union forms the candidate set; the intersection earns a candidate-quality boost.
5. 5

   Multi-signal ranking

   The candidate set is scored against the four pillars (relevance, distance, prominence, behavioural). The contribution of each pillar to the final score is query-dependent: short-distance queries weight distance more, long-tail specialist queries weight relevance more.
6. 6

   Filters and de-duplication

   The Possum filter de-duplicates listings sharing addresses or sharing very similar names and categories within close proximity. The spam filter removes obviously fake or policy-violating listings. The quality filter demotes incomplete or stale listings. Personalisation then re-orders based on the searcher's history.
7. 7

   Surface routing

   The top results are placed into one of several surfaces: the SERP Local Pack (typically three results), the Local Finder (when "more places" is clicked), the Maps app, an AI Overview with local intent, or AI Mode's conversational response. The ranking is similar across surfaces but not identical; each surface has its own re-ranking step.

The ranking factor stack

Pulling together a decade of independent ranking-factor surveys, the observational pattern across our own customer base, and Google's own documentation, the approximate weight of each signal category on Map Pack ranking in July 2026 looks like this:

Pillar 1: Relevance, in depth

Relevance is how the algorithm decides whether your listing should be a candidate for a specific query at all, and Google's category guidance is where candidate-set membership starts.⁷ The signal stack inside relevance, in approximate descending order of impact:

Pillar 2: Distance, in depth

Distance is the most-misunderstood pillar because the phrase implies a simple straight-line measurement. In practice, distance is a relevance-weighted radius modified by user signals and query intent. Four mechanics are worth understanding:

Pillar 3: Prominence, in depth

Prominence is how well-known your business is to Google and to its users. It is the slowest pillar to move and the one with the highest ceiling. The signal stack:

Pillar 4: Behavioural signals, the silent accelerant

Behavioural signals are the part of the local algorithm that has grown the most in the past three years. Google does not name them in its public documentation, but the patterns are visible in any decent before-and-after testing across listings with otherwise-identical signal stacks. The behavioural signals the algorithm reads:

These signals compound with the other three pillars. A high-CTR listing for a category-relevant query gets a relevance bump above what its raw signal stack would predict. A high-direction-request listing gets a prominence bump. The compounding is the reason listings that "shouldn't" rank sometimes do, and listings that "should" rank sometimes don't.

Behavioural signals have become materially more influential since 2022. Time and again, a listing with strong engagement signals outranks a listing with stronger raw signals but weak engagement. It is one of the clearest patterns we see, and one of the most under-optimised.

Filters Google applies after ranking

After candidates are scored, several filters operate on the ranked list before it is shown to the user. These filters are the cause of most "we should be ranking but we aren't" diagnoses we run for agencies:

• Possum filter (September 2016, modified by Hawk in 2017). De-duplicates listings sharing the same address or very similar names and categories within close proximity. After Hawk, the filter only triggers when the listings are very close together; mid-distance siblings are no longer always filtered. The most common cause of unexpected ranking absence in multi-location businesses with multiple locations in the same building or business park.
• Vicinity filter (December 2021). Reduced ability of distant businesses with keyword-stuffed names to rank for proximity queries. The filter applies a sharp distance falloff past a category-dependent radius.
• Spam filter (continuous). Fake businesses, obvious lead-generation listings, and listings flagged through the report-a-problem flow. Suspension is an upstream binary, not a filter.
• Quality filter (continuous). Incomplete or stale listings are demoted within candidate sets. Listings with no photos, no description, no services listed, and no recent posts can rank, but they will lose head-to-head to a more complete peer with similar raw signals.
• Personalisation re-ranking (continuous). Search history, past clicks for similar queries, and preferred businesses re-order results for the individual searcher. This is why two people standing next to each other can see different Map Packs for the same query.

The surfaces: where local results actually appear

The same underlying algorithm produces results across several different surfaces. The ranking is similar across surfaces but not identical; each surface has its own re-ranking step and its own display constraints:

Algorithm history: the named updates that matter

Local algorithm changes happen continuously, but a handful of named updates produced step-change effects that still shape how the system behaves today. The core and spam update names are Google's own and appear on its ranking updates record.⁸ Possum, Hawk, Bedlam and Vicinity are community names: Google confirmed the changes behind Bedlam and Vicinity, but never confirmed Possum or Hawk.

AI search and local in 2026

The local algorithm now feeds both traditional surfaces (Map Pack, Local Finder, Maps app) and the AI surfaces (AI Overviews, AI Mode, and through retrieval, Claude, ChatGPT, and Perplexity). The mechanics differ in subtle but important ways:

Across third-party AI assistants, the pattern in 2026 is that each vendor uses its own mix: web search providers (OpenAI says ChatGPT uses third-party search providers; Claude's web search runs on Brave Search), licensed partner data such as the Yelp reviews and listings now in ChatGPT, and their own crawlers. None documents using Google's local index. Pages associated with a clean entity record tend to be cited disproportionately, even when their raw rankings are middling.

One newer Google surface sits alongside these. Ask Maps, launched in March 2026 in the US and India, answers conversational questions inside Google Maps using Gemini.⁹ Google says it draws on over 300 million places and reviews from more than 500 million contributors, and adjusts for places you have searched for or saved. Treat it like AI Mode: complete profile fields and review depth are what it has to work with.

Common myths and what is actually true

How to test and instrument

The local algorithm responds to changes on timescales ranging from hours (some behavioural and proximity changes) to weeks (review velocity changes, content updates) to months (link and citation changes, brand-mention compounding). Reliable instrumentation is the difference between knowing what is working and guessing:

• Geo-grid rank tracking. Measure rank for your target queries at multiple geo-points around your service area, not just a single point. Rankings vary continuously across metres of geography; a single-point average is misleading. Our Geo-Grid Rank Tracking feature is purpose-built for this.
• Mobile and desktop separately. Proximity weighting differs between mobile (GPS-precise) and desktop (IP-coarse). Track both for any geography you care about.
• Before-and-after testing on single levers. Change one thing at a time, wait for the algorithm to absorb the change (a week is usually enough for GBP-field changes), then measure delta. Multi-lever changes are diagnostically useless.
• Behavioural metrics from GBP Performance. Pull the Performance API into a warehouse and watch impressions, calls, direction requests, and website clicks over time. The API has no click-through metric, so work out actions per impression yourself. These leading indicators predict ranking changes before rank tracking catches them.
• Cross-platform monitoring. Map Pack rank is one number. Total business visibility is rank across the Map Pack, Local Finder, Maps, AI Overview citations, and AI Mode mentions. The same listing optimisations move all of them, but not at the same speed.

The audit checklist

• Primary GBP category is the narrowest accurate option Google offers; review against the full searchable list
• Three to six secondary categories that each genuinely describe additional work you do
• Services list populated with 10 to 30 entries, each with a brief description
• Every applicable attribute ticked; reviewed at least quarterly
• 20+ recent photos across exterior, interior, team, products, and work-in-progress
• Description uses the full 750 characters, leads with what you do and who you serve
• Every review from the last quarter has a reply within 48 hours of being posted
• The top 10 questions a customer would ask are answered somewhere Google can read: services, attributes, or an FAQ page on the site
• Posts published within the last four weeks
• Special hours scheduled for upcoming public holidays and closures
• Website has LocalBusiness schema markup with consistent NAP
• sameAs links from authoritative identifiers (Companies House, regulator IDs, Wikidata if applicable)
• Citation footprint consistent across the high-trust local directories for your country
• Backlinks from at least a handful of local publications, partners, or industry bodies
• Geo-grid rank tracking running for your top three commercial queries
• GBP Performance API or Insights reviewed monthly for behavioural-signal trend changes
• Mobile and desktop ranking tracked separately for the same queries
• AI Overview and AI Mode visibility checked for your top three queries each month

Where the algorithm is heading

Looking at the trajectory of the past four updates, the practical expectations for the rest of 2026 and into 2027 are:

• Continued weight shift toward behavioural signals. Every update since 2022 has nudged behavioural weight higher. Expect this to continue.
• Tighter entity-record requirements. Listings without clean schema, sameAs convergence, and consistent NAP will increasingly underperform listings with the same raw GBP signals but cleaner entity records.
• AI surfaces becoming a larger share of impression counts. AI Overviews and AI Mode are expanding their query coverage; the Map Pack remains the dominant surface but its share of total local impressions has declined.
• Reviews continuing to gain weight, but conditioned on authenticity. Review spam detection has tightened materially in 2024 and 2025. Volume and velocity matter, but only if the reviews look organic.
• Lower returns from citation-volume tactics. Building citations on long-tail directories has been losing weight for years. The trend is continuing.

Where to go next