Mobile, Local and Voice Search

By Bill Hartzer · Updated 19 August 2026

Why these three belong together

Mobile, local and voice search are usually treated as three subjects. They are really three consequences of one event: between roughly 2007 and 2015 the device people searched from stopped being a computer on a desk and became a phone in a pocket.

A phone differs from a desktop computer in three ways that matter to a search engine, and each produced a different structural change.

It knows where it is, to within a few metres, without being asked. That made location an implicit part of almost every query and forced engines to build and rank a second kind of index — one full of businesses and places rather than web pages.

It has a microphone and no keyboard worth using, which made spoken queries practical and changed the queries themselves: longer, phrased as questions, full of the small grammatical words keyword matching had always discarded.

And it has a small screen, sometimes none at all. That compressed the results page and eventually removed it, because a device answering aloud cannot read out ten links. The single-answer format voice required is the format AI search now delivers by default.

None of this was a redesign. It was an engine built for one device, renovated over fifteen years around another.

Mobile: from a shrunk desktop to the primary index

The first phase was cosmetic. Engines detected a mobile browser and served a narrower page. The web itself, and the index behind it, remained a desktop artefact.

The first ranking consequence came on 21 April 2015, when Google made mobile usability a ranking input for searches performed on mobile devices. The industry nicknamed it Mobilegeddon before it happened; the anticipation exceeded the effect and the name outlived both. Around the same period Google stated that more searches were taking place on mobile devices than on computers in the United States and several other countries — quoted constantly ever since, and never updated with a published global split.

The deeper change was mobile-first indexing, and it took seven years. Google announced the concept in November 2016, began rolling it out in March 2018, made it the default for newly discovered sites in May 2019, announced it for the whole web in March 2020 with a deadline later extended, published a progress update in November 2021, and declared the transition finished in a Search Central post, "Mobile-first indexing has landed", in October 2023.

What it means is worth stating precisely, because it is often described as a preference when it is an architecture. Google's index is built from what its smartphone crawler retrieves and renders. Where the two versions of a site differ — in text, in structured data, in which images and links are present — the mobile version is the one that exists as far as the index is concerned. Mobile stopped being a mode of viewing search results and became the definition of the corpus.

Local: the index that is not a crawl

Local search is the one part of a general search engine where the underlying data does not come from crawling the web at all, and almost nothing written about it makes that clear.

When a query produces a map with a short list of businesses beneath it — the "local pack" — those entries are not web pages that were crawled and ranked. They are records in a business directory: name, category, address, coordinates, opening hours, phone number, photographs, reviews. The bulk of that data is submitted by the businesses themselves and verified by the platform, historically by post, telephone or video. A crawler discovers pages; a local index accepts submissions.

Google's local infrastructure was assembled quickly. It acquired the Sydney company Where 2 Technologies in October 2004, turned its desktop mapping application into a web application, and announced Google Maps on 8 February 2005. Google Local for Mobile followed in beta that November, Maps was preloaded on the first iPhone in 2007, and turn-by-turn navigation arrived as an Android beta in October 2009.

The ranking side moved in parallel. In February 2012 the change known as Venice began localising ordinary web results — different results in different cities for queries containing no place name at all, on the inference that "plumber" means a nearby plumber. On 24 July 2014 the update known as Pigeon tied local ranking more tightly to the general web ranking systems, launching in US English across both Search and Maps. Around 2015 the local pack was widely reported to have been cut from seven entries to three; Google did not announce it.

Google's public description of local ranking rests on three components — how well a listing matches the query, how far it is from the searcher, and how prominent the business is — useful shorthand because distance has no equivalent anywhere else in web ranking. It is also why local search has an integrity problem ordinary web search does not: listings are self-reported, so fake and duplicated entries are a persistent category of abuse, and the countermeasure is verification rather than crawling.

Voice: the microphone as a query interface

Voice search predates the smartphone assistant by several years, and its first widely used implementation was, quite openly, a data-collection exercise.

GOOG-411 launched in 2007 as a free telephone directory service in the United States and Canada: dial a toll-free number, say a business and a city, get connected. Google was explicit that the point was to build a large phoneme database from real users' voice queries to improve its speech recognition. It was discontinued on 12 November 2010, widely read at the time as meaning Google had collected enough speech.

The rest of the timeline is a sequence of interfaces rather than algorithms:

November 2008 — voice search added to the Google Mobile App on iPhone.
June 2011 — voice search on the desktop, initially only in Google Chrome.
October 2011 — Siri ships with the iPhone 4S, making a voice assistant a headline feature of a mainstream phone for the first time.
2012 — Google Now, a predictive card-based assistant, absorbs voice actions on Android.
2014 — Microsoft's Cortana arrives on Windows Phone; Amazon's Echo and Alexa put a voice interface in the home with no screen at all.
18 May 2016 — Google Assistant is unveiled at Google's developer conference.
14 March 2025 — Google announces that Assistant will stop working on most devices, replaced by Gemini, with the older Assistant surviving only on lower-end and legacy Android hardware.

That last date is the tidiest possible epitaph for the category. The voice assistant did not win or lose; it was absorbed into a language model.

What voice actually changed about search

Two things, and only one of them gets discussed.

The first is the shape of the query. Typed queries are telegraphic — people drop articles, prepositions and question words because typing costs effort and a decade of keyword search taught them those words were ignored anyway. Spoken queries are not compressed. People speak in sentences and use exactly the small grammatical words keyword matching throws away. Google's Hummingbird rewrite of August 2013 arrived in the middle of this shift, and the neural systems that followed — RankBrain from 2015, BERT from October 2019 — are in large part systems for recovering the meaning those words carry.

The second, and more consequential, is the shape of the answer. A results page is parallel: ten results offered simultaneously, with the choice left to the reader. Speech is serial. A device answering aloud reads one answer, and the ninth-best result does not exist. Voice therefore forced search engines to commit — to select a single passage from a single source and present it as the answer.

That is the same architecture as the featured snippet, and the direct ancestor of the AI-generated answer now sitting at the top of a large share of results pages. The publisher-traffic argument of the mid-2020s — that answering the question on the results page removes the reason to visit the source — was fully present in voice search a decade earlier, in a form nobody complained about because the volumes were small.

Voice also skews the query mix toward things awkward to type and immediate to need: directions, opening hours, nearby businesses, quick facts. This is why voice and local are hard to separate — many voice queries are local queries, answered from the business directory rather than the web index.

The statistic that will not die

Any search for the size of voice search will produce the claim that half of all searches would be voice by 2020. It is repeated in trade press, conference slides and AI-generated summaries, usually with a market research firm's name attached and no link.

The claim appears to descend from a 2014 prediction by Andrew Ng, then chief scientist at Baidu, that at least half of all searches would be images or speech by 2020 — a different proposition, including visual search, and an informal forecast rather than a measurement. The version in circulation drops the images, converts a prediction into a finding, and cites a source that does not appear to have published it.

The honest position is that no search engine publishes voice query volumes. Google, Microsoft, Apple and Amazon have released selective figures at product launches, none of which constitutes a trackable series. Any percentage attributed to voice in 2026 should be treated as unsourced unless it comes with a named study, a sample and a date. It is a clean example of how a plausible figure with no origin becomes a permanent fact.

Where the three ended up

By 2026 none of the three is a separate product, which is why writing about them separately has become misleading.

Mobile is no longer a variant of search; it is the index, with no desktop corpus to fall back to. Local is both a vertical inside the general engine and a destination in its own right, and remains the one part of the system built on submitted rather than crawled data — the part where verification rather than ranking does most of the quality work. Voice has been absorbed into the assistant, and the assistant into a language model. Google's replacement of Assistant with Gemini from March 2025 is the clearest marker, but the pattern holds across the industry: speech is an input method feeding a conversational system, not a distinct search product with its own results.

The pattern is not universal. In South Korea, Naver's results page has always been built from its own curated content blocks rather than a list of web links, so the mobile transition looked quite different there. In China, Baidu's mobile and voice products developed inside an app ecosystem rather than around a browser. Microsoft's Cortana ended as a Windows feature, and Bing's voice work now surfaces through Copilot. "How search changed on mobile" is largely a description of how Google changed — a statement about market share as much as technology.

Frequently asked questions

What is mobile-first indexing?

It means Google builds its index from what its smartphone crawler retrieves and renders, rather than from the desktop version of a page. Where the two versions of a site differ, the mobile one is the version that exists as far as the index is concerned. Google announced the approach in November 2016, began rolling it out in March 2018, and declared the transition complete in October 2023.

When did mobile searches overtake desktop?

Google stated in 2015 that more searches were happening on mobile devices than on computers in the United States and several other countries. That statement is the origin of almost every citation on the subject, and Google has not published an updated global split since. Treat any precise current percentage as an estimate from a measurement firm rather than a figure from a search engine.

Where do local search results come from?

From a business directory, not from a web crawl. The entries in a local pack are records containing a name, category, address, coordinates, hours, phone number, photographs and reviews, largely submitted by the businesses themselves and then verified by the platform. This is the one part of a general search engine whose primary data source is submission rather than crawling.

What was the Google Pigeon update?

Pigeon was a local search ranking change launched on 24 July 2014, initially in US English, which tied local ranking more closely to Google's general web ranking systems and affected both Search and Maps. It followed the February 2012 change known as Venice, which began localising ordinary web results for queries that contained no place name at all.

What was GOOG-411?

A free telephone directory service Google ran in the United States and Canada from 2007: callers dialled a toll-free number, spoke a business and a city, and were connected. Google stated its purpose was to build a large phoneme database from real voice queries to improve its speech recognition. It was discontinued on 12 November 2010, widely read as meaning Google had collected enough speech.

How did voice search change search results?

It forced engines to commit to one answer. A results page offers ten results in parallel and lets the reader choose; a device answering aloud must read a single passage from a single source. That single-answer format is the same architecture as the featured snippet and the direct ancestor of AI-generated answers, along with the publisher-traffic argument that came with them.

Is it true that half of all searches are voice searches?

There is no evidence for it. The claim descends from a 2014 prediction by Andrew Ng, then at Baidu, that at least half of searches would be images or speech by 2020 — a forecast about a broader category, later repeated as a finding with a source attached that does not appear to have published it. No search engine publishes voice query volumes.

Does Google Assistant still exist?

Only in a residual form. Google announced on 14 March 2025 that Assistant would stop working on most devices, replaced by Gemini, with the older Assistant remaining on lower-end and legacy Android hardware. The move is representative rather than unusual: across the industry, voice has become an input to a conversational language model rather than a search product in its own right.

Mobile, local and voice search