Google Ads deep dive

To select an ad, Google Ads needs to understand the offer on the page. The app’s code supports several steps: fetch the page, check that it is accessible, store its content in a database used by Ads, then prepare it for AI.

Text, vectors and estimates such as click probability are produced on Google’s servers. Vectors represent content as numbers so its meaning can be compared with other content. The app reveals the information sent to Google, the results expected and the checks planned. It does not reveal the full workings of the AI models.

Fetch pages and check access

Several functions in the code handle crawling, updating page content and checking URLs.

Google Ads can fetch page content, check that it is usable and update it. The code also includes a landing page content database, named LANDING_PAGE_REPOSITORY, and tracks whether a page is available for asset creation.

The ASSET_GENERATION_STATUS_URL_NOT_INDEXED message indicates that a page is missing from the database used by Ads. It does not describe its presence in Google’s organic search results. A page can therefore be stored in one system without being stored in the other, even if the two share some data.

Find useful information on the page

The code supports extracting different parts of the page separately.

Google Ads distinguishes the title, description, sections, important passages, text snippets, structured data and images. Each part can serve a different purpose.

The title and business name can help identify the offer. Page text and existing assets can help write an ad. Short versions of pages make it easier to compare them when choosing a URL. The system can also look for the passage that supports an ad claim: this check is called grounding.

Structured data describes the business, products, categories and their attributes in a machine-readable format. Visible text provides explanations and evidence. The two need to be consistent: structured data does not replace page content.

Prepare different versions of the content for AI

The code supports several ways to prepare content: combine the business name, title, meta tags, description lists and page text, then shorten the result. Google Ads can therefore create different versions of the content from the same HTML.

These versions help find information, compare pages, give instructions to AI and check the text it produces. Some use Markdown, a simple text format that preserves headings and lists. The purpose is practical: understand the offer, choose the right page and write an ad that reflects its content.

From page content to ad text

The code names the main preparation and checking operations in TransformationName_Enum.

The system gathers information from the page and other sources, then prepares instructions for the AI. It checks the generated text and records which information was used. The list of operations helps reconstruct this process, but does not establish the exact order of every step on its own.

A reference profile for the business

Names in the code point to a business profile separate from its web pages. It can bring together the business’s identity, category, description, brand characteristics and several content sources.

The AI uses this profile and its knowledge of the products and brand to assess what the advertiser is trying to promote.

For marketing teams, these sources need to tell the same story. Different categories, contradictory descriptions or inconsistent offers make it harder to create ads and match them to the right queries.

Google Ads knows page types, not just URLs

To create an asset, the system can draw on several website pages and other sources. The AdAssistantAssetExtractionSourcePB_Enum list distinguishes pages by their purpose.

The home, About, Contact and Why-us pages are named separately in the code: HOME_PAGE, ABOUT_PAGE, CONTACT_PAGE and WHY_PAGE. They can therefore provide information for creating ads, alongside the product page that receives the click.

The code tracks each source separately for asset suggestions: the landing page, recent assets, past campaigns, stock images, social networks, Merchant Center or Business Profile. AssetAutomationSuggestionsStatus defines 17 tracking fields, including smartGenaiAssetSuggestionStatus, behind what the interface presents as a single feature.

Which queries help prepare the ads?

The name LP_GROUNDED indicates that the query or keyword is compared with the landing page content. QuerySpec settings define the query source, number of queries, time period and a maximum per ad group. Only part of the available history is therefore selected to prepare ad text.

References to similar queries, synonyms, DSA and broad match concern finding more relevant queries. SQM_ORGANIC_TRAFFIC names a variant using organic traffic information. Combining paid and organic data can support reporting and suggestions; it does not establish that an SEO score enters the Ad Rank calculation.

The code distinguishes matches based on search context, close variants and matches associated with Performance Max. Reports also retain the match source: AI Max without keywords or AI Max broad match. A dedicated report links each search term to the landing page and ad headline that were served.

Public priority rules favour a Search campaign when a query is identical to an eligible keyword. Search Themes have a priority comparable to phrase and broad keywords. The keywordless systems of AI Max, DSA and Performance Max cover the remaining demand.

Do Search Themes bring additional traffic?

Public documentation describes measuring whether a Search Theme brings traffic that Performance Max would not have found on its own. SEARCH_SHADOW most likely refers to that baseline. A rise in reported conversions is therefore not enough: check whether the theme reaches new searches or takes credit for ones already covered.

NavBoost and QBST: queries that help write ads

The code defines several versions of the Coauthor writing functions. Five functions each appear in three forms: without an extra label, with NavBoost disabled (NAVBOOST_DISABLED) and with filtering for rare queries (K_ANON_NAVBOOST). The table preserves their technical IDs.

These names appear among functions that prepare landing pages, queries and ad headlines. They indicate that queries linked to a page can help write ad text.

The antitrust trial describes NavBoost as a system that links queries and documents from memorised click data. A Google patent with a 2004 priority date describes the reverse move: starting from a document, retrieve the queries that led to clicks, then use them as keywords or creative material. The K_ANON_NAVBOOST variant points to filtering meant to keep aggregate behaviour while dropping queries that are too rare.

QBST stands for Query Based Salient Terms. It identifies important words and phrases on pages relevant to a query, also using click data. The names QBST_TEXT_ASSET_SUGGESTION and QBST_ADVERTISER_TEXT_ASSET in WorkflowAssetGenerationKeywordSourcePB_Enum identify it as a source of terms for asset creation.

From website information to ad selection

Names grouped under IDs 226 to 255 in SignalName_Enum refer to the page, reports, text creation, evaluation, selection and the brief. The steps below offer an interpretation.

The three AI Max features

AI Max for Search was announced in beta on 6 May 2025 and left beta on 15 April 2026. It bundles three features.

The switch is under way. Since 1 September 2026, campaigns that used automatically created assets (ACA) are moved to AI Max automatically with search term matching and text customization switched on. Campaigns that used campaign-level broad match receive only search term matching. DSA campaigns will follow from February 2027, with all three features and their URL controls preserved.

The app includes AI Max settings at campaign and ad group level, plus URL inclusions and exclusions. The ai_max_search_term_ad_combination_view report links a search term to the landing page and ad headline that were served.

Final URL expansion uses the site architecture

A page that is hard to find, poorly linked or unclear about its offer is harder to select. A well-structured category, service or service-area page can answer a search more precisely than a generic page.

AI Brief: give instructions in your own words

AI Brief, announced on 30 April 2026, organises advertiser instructions into three categories: messaging, query matching and audiences. Names in the code refer to these same uses.

The brief sets rules for choosing queries, writing text and selecting ads. It complements keywords, exclusions, bidding goals and brand guidelines.

What you give the AI changes the result

The code can combine the page title and description, keywords, existing assets, business information and queries. It also includes writing instructions, examples and a response format. Together, these form a ‘recipe’ for creating ads.

The code includes several systems for running models, with different versions and settings. The choice can reflect language, speed, availability or cost needs. Sax refers to Saxml, software published by Google. Beyond, Evergreen, Servo, Steelmill and Uniserve remain internal names whose precise roles are not publicly documented.

Six years of changes in ad creation

The AssetSelectionVersionPB_Enum list contains 295 versions of methods for creating and selecting assets. Of these, 157 carry a date, from May 2020 to May 2026. They show how the techniques defined in the code developed over time. Their presence does not reveal which are active on an account today.

Seven categories of ad checks

The code contains two lists for evaluating assets: 88 score types in AssetScore_ScoreType and 60 quality indicators in AssetQualitySignalType_Enum. They share 59 names but remain separate lists. Many of the checks assess whether ad text reflects the page content and whether its claims are supported.

Some scores assess relevance, expected clicks or advice for improving headlines. Others try to identify the page sentences used to create the text. These checks answer several questions: is the ad relevant, appealing and consistent with its source content?

These checks can be grouped into seven categories. They help select text, reject problematic content and provide feedback to the advertiser. They are not all applied to every asset.

The checks depend on the text source, model, language, asset type and features enabled for the account. All 88 score types are therefore not calculated for every proposed ad.

Check each claim against the page

Settings allow different checks for a headline, description, product or keyword. The system can compare a whole text or specific words with the page. It can also measure how much of the text is supported by a source and detect negations that change the meaning.

The code can retain the exact passages supporting part of an ad’s text, along with their language and limits on length and number. This is the grounding check introduced earlier.

Which information helps verify ad claims?

The AssetQualitySignalType_Enum list defines 38 indicators related to consistency with source content. The variants mainly change the information provided to check the text: the title alone, description, business name, snippets or full page content.

These variants address practical questions: is the title enough to verify a claim? Does the meta description help identify the right source? Does the business name prevent confusion between brands? Does the full text add useful information or make the check harder?

Other checks look for likely source sentences, assess accuracy and usefulness, or identify contradictions and changes in meaning or tone.

The code also names an internal test set with reference answers for assessing these checks. Some settings can return only claims whose source could not be confirmed.

Past results help create the next texts

The system can also use results from ads already served. The SignalName_Enum list includes several measures of their impressions and clicks.

Texts that earned clicks can provide information or examples for creating the next ones. Some method names also refer to objectives for click probability (pCTR), landing page experience (pLQ) or both (PCTR_PLQ). These estimates can therefore guide text creation as well as play a role in the auction.

For paid search teams, current assets can therefore influence future suggestions. Texts with little variety give the AI few different examples. A headline that attracts clicks with an unsupported promise may look good for CTR while creating a poor experience after the click.

Several selections happen before Ad Rank

Before the auction, Google Ads has already selected usable pages, information for writing text, relevant queries and asset combinations. It may also have selected products, images or videos. Ad Rank comes after these choices.

The 88 asset score types help evaluate text as it is prepared. They do not form a second Ad Rank. The website, product catalogue, structured data and brand guidelines therefore play a role earlier in the process: they supply the information the system can choose from.

The landing page plays a role at several points: its content helps create the ad, its expected quality matters in the auction, and diagnostics help the advertiser improve it. It does not determine ranking or click price on its own.

Ad Mixer: rank ads and account for the long term

According to a Google document provided to the CMA, the UK competition authority, a few hundred relevant ads reach Ad Mixer. This system assigns them an Ad Score or LTV Score. Trial documents connect this score with the public name Ad Rank.

Other trial documents explain that Google accounts for expected ad quality (pCQ) and landing page quality (pLQ). The calculation combines expected click revenue and the negative effects of a poor experience. It also accounts for the risk that users click ads less often in the future, known as ads blindness.

Document UPX0010 explicitly links LTV Score and Ad Rank. Click price also depends on pricing mechanisms separate from ad quality. The record names squashing, format pricing and rGSP, detailed below. Judge Mehta’s opinion states that Google used them to increase text ad prices.

Patents filed as early as 2005 already distinguish click probability from landing page quality. The code also names text creation methods guided by pCTR, pLQ or both. For paid search teams, the implication is clear: a headline can attract clicks while creating an expectation the page does not meet.

Ad Rank thresholds determine whether an ad can occupy a placement. They vary with quality, context, position and the search topic. Even without a direct competitor, they can impose a minimum price. Other pricing settings can also change CPC.

Google also ranks its advice

For each recommendation, the code defines one effect on Optimization Score if it is applied and another if it is dismissed. These effects can be tracked at account and campaign level. The score therefore also depends on how you handle recommendations.

The budget is steered automatically too

The code describes a temporary budget increase with a possible return to the original budget. It specifies the information needed to track this adjustment.

AdaptiveBudgetsSettings also defines limits for automatic adjustments: an allowed percentage, a minimum and a maximum. Campaign-specific settings and an opt-out are included. This allows the system to limit the size of budget changes.

Google states that for ads served in AI Overviews two contexts are taken into account: the user query and the content of the AI Overview. The Search or Shopping ad has to win the auction, answer the query and stay relevant to the information in the generated answer.

The code contains functions that could help match ads to conversation content.

In a classic search, Google connects the query, ad and landing page. With an AI-generated answer, the answer’s content also matters. Several 2023 patents describe ads adapted to conversations: identify important terms in exchanges, clarify the need across messages and adjust the ad text. These patents describe possibilities, not proof of deployment.

Shopping also uses the product feed and website

AI Max for Shopping uses Merchant Center feed data, site content and title generation to answer more conversational queries. Final URL expansion can walk the site to identify categories, new arrivals, commercial editorial pages and other relevant destinations.

The code includes product information (gmc_product_info), Merchant Center asset suggestions, information about titles and descriptions, and landing page variants. The productFidelityScore assesses whether a generated asset describes the product accurately.

Assistants for preparing and editing campaigns

The name GENERIC_MUTATE indicates that an assistant can prepare or apply account changes, subject to the product’s controls. Another set of functions, Ads Guide, brings together conversation history, website information, personalisation and campaign creation. It also includes specialised capabilities and models able to reason through several steps.

The products Google has announced point the same way: Ads Advisor, Analytics Advisor and Ask Advisor. Google states that approval is requested before an action is applied.

To manage these assistants, instructions and sources need to be consistent. A brand rule is hard to apply if the website, product feed, Business Profile and assets give conflicting information. The assistant works with the data it can access.

We analysed app versions 3.34 and 3.37 to identify supported functions, exchanged information and available settings. Both versions were compared on the same architecture to distinguish version changes from differences in how the app was built. This work was complemented by observing the ads.google.com web interface: several traffic captures from real sessions were used to cross-check the site’s vocabulary and schemas against those found in the app.

A tool developed for the study read the program’s structure beyond a simple word search. Checks confirmed that the analysis covered the expected data: 3,018 groups read out of 3,018, followed by 220 matching fields out of 225 when compared with the web interface. The version 3.37 inventory includes 1,638 technical value lists and 39,409 entries. The detailed method is available below.

This interpretation is compared with US v. Google trial evidence, product documentation, Google Research publications and the patents listed below. These sources help interpret the functions found in the code.

The app does not reveal internal model settings, decision thresholds or the features actually enabled on each account. Some connections remain hypotheses, including the role of SEARCH_SHADOW and the order of checks or AI Max steps. Beyond, Evergreen, Servo, Steelmill, Uniserve and Sequoia remain internal names without public documentation.

Main sources

US v. Google trial

• UPX0010, Google’s response to the CMA on Search ad ranking and pricing (2020); note 30: pCTR uses no landing page information; note 34: pCTR, pCQ and pLQ are used in the auction
• UPX0008, Introduction to the Search Ads Auction
• Plaintiffs' Proposed Findings of Fact, ECF 839 (pCQ, click cost, LTV)
• Judge Mehta's opinion, 5 August 2024 (LTV formula, squashing, rGSP)

Google documentation and announcements

• Google special-case crawlers and AdsBot
• AI Max for Search, beta announcement, 6 May 2025
• AI Max migration: distinct settings for broad match and ACA campaigns (Google Ads Developer Blog, 12 August 2026)
• AI Max out of beta, September 2026 auto-upgrade and DSA sunset pushed to February 2027
• AI Brief, announcement of 30 April 2026
• AI Max, product documentation
• Ads in AI Overviews
• Search ads and landing page navigation, 5 February 2025
• RSA guide 2023: how Google assembles and scores combinations
• Ad Strength for RSAs
• Quality Score and components
• Optimization Score and dismissed recommendations

Research publications

• Focusing on the Long-term, KDD 2015 (ads blindness)
• On the Factory Floor, 2022 (pCTR model)
• Advertiser Content Understanding via LLMs for Google Ads Safety, CIKM 2024
• Scaling Up LLM Reviews for Google Ads Content Moderation, WSDM 2024
• Attribute First, then Generate, Slobodkin et al. 2024 (methodological convergence, no mention of ads)

Patents

• US20050222901A1 (2004), queries that led to clicks on a document, supplied as keywords
• US20070156621A1 (2005), post-click quality parameter separate from CTR
• US10963916B2, Ad Strength: ensemble score of an asset group
• WO2024243265A1, checking claims against sources, with an acceptance threshold and replacement
• EP4565990A1, US20250148364A1, US20250086434A1, ads in conversational answers