Ignacio Cánovas — AI Product Strategist

01 · Context

Two types of searchers. Which one are you?

Not everyone uses Chrome the same way. Some users open a new tab and start from Google's homepage. Others never see that page — they go straight to the URL bar. Two types of Chrome users. Two parallel studies screened for both intentionally.

NTP participants were confirmed homepage users on Mac/macOS Sonoma. Omnibox participants were confirmed URL-bar-first users on the same setup. Both desktop only, three weeks each.

Study 01

New Tab Page

Homepage users — the most iconic search interface in the world.

Study 02

Omnibox

URL-bar-first users who bypass the homepage entirely.

"Two studies. Three weeks each. One researcher. Synthesized with DMP — built to make AI-assisted research defensible at scale."

02 · Methodology

How observations became defensible findings

Every finding in this case study is traceable back to a specific observation, in a specific session, from a specific participant. The pipeline that made this possible is DMP — Data Mise en Place.

Raw Input

Every session: transcript + observation notes. When a user says one thing and does another, that divergence is itself a finding.

Clean Notes

AI-assisted normalization. The agent prepares the data. It does not analyze it.

Atomic Observations

Single, discrete, irreducible moments. One observation, one thing that happened.

Evidence Type Classification

Action/Behavior, Perception, Value, Trust, Intent, Outcome, or Friction.

UX Area Tagging

Evidence type = what kind of truth. UX area = where it lives. Together: every finding is typed and located.

Evidence Types

Action / Behavior

What the user did. When action and narration diverge, action takes precedence.

Perception

What the user thought was happening. Misperception is often the root cause of friction.

Value

What the user found useful or worth using again.

Trust

What built or eroded confidence across interactions.

Intent

What the user was trying to accomplish.

Outcome

What actually happened. Closes the loop on intent and action.

Friction

Where the interaction broke down. Distinct from outcome: a user can complete a task while experiencing significant friction.

03 · AIM Button

Finding the button was just the beginning

Most users landed on the SRP without activating AI Mode — habit took them straight past the button. But those who discovered AIM adopted it quickly. What they never learned was whether it was on or off.

First encounter: missed by habit

The AIM button was not hidden — it was outside the field of attention at the moment that mattered. Discovery was accidental, not deliberate.

Adopted quickly, state never clear

Once found, users adopted fast. The problem: state visibility. Users never knew if AI Mode was on or off — some clicked multiple times cycling through states.

Two paths into a heavier layer — no warning for either

Deep Dive Search (inside AIO) and Deep Search (inside AIM) could both trigger 4–5+ minute waits. No cue. No time estimate. No way to stop.

Trust erosion: latency without context

"Thinking..." told users the system was busy — nothing else. Many concluded it had broken. Trust slipped not because AI failed to deliver, but because it failed to communicate.

Spatial friction: far from where searching starts

After adopting the button, users noticed how far it sat from the input field. Not a blocker — a persistent friction point even for users who had fully learned the interface.

04 · Entry Points

Where users started shaped their first impression. Nothing more.

The entry point mattered at the beginning. After that, the experience was the same for both groups.

NTP Users

Gradual dissonance

Familiar at first. Day 4 breaking point: customization options revealed the full scope of changes. The NTP stopped feeling like Google.

Omnibox Users

Task-first arrival

The URL bar is functional, not branded. AIM button easier to miss. Both groups then landed on the same SRP.

05 · SRP Experience

Once they landed on the SRP, the mental model converged

Regardless of entry point, both groups encountered the same layout, the same AI layers, and the same chips. What they valued: conversational answers, the sourcing sidebar (once found, strongly positive — users wanted inline citations), follow-up questions, and AI available without leaving Chrome.

"Users didn't distrust the AI. They distrusted the interface around it. The capability earned trust every time it was reached. The path to it didn't."

06 · Where It Broke

When search behavior collides with AI chat interface

Twenty years of muscle memory: type at the top and hit Enter. Once inside AIM fulfillment, that instinct ran into a fundamentally different interaction model — two input fields, two paradigms, one unresolved interface.

Behavioral conflict

Twenty years: type at the top. AI chat: type at the bottom. When in doubt, users went back to the top — and couldn't do anything from there.

Realbox — read only

The Realbox showed the original query but couldn't be edited in AI Mode. Users instinctively tried to refine from there — and hit a wall.

Visual communication failures — systemic

No activation feedback when clicking. "Thinking..." communicated busy but not progress or duration. AIM / AIO / Research a Topic redundancy never resolved.

07 · Action Chips

Great shortcuts when they worked. A dealbreaker when slow.

Research a Topic

Praised, then abandoned

Clarifying questions: most praised interaction. Processing time 4–5+ min with no way to stop. Users felt trapped and left.

Ask Google about a Tab

Concept valued, execution limited

Only referenced the last tab — users had to reorganize tabs to use it on others.

Create an Image

Excited, then broken

Expected a generated image, received search results. Activation/deactivation loop confused users.

Two entry points. One set of problems at the destination.

Most users never activated AI Mode

Invisibility, unreadable state, spatial disconnect — all equal across both groups. Discovery was accidental.

The toggle problem is a trust problem

Without clear state feedback, users couldn't trust their actions or the output — even when it was good.

Search and chat: two paradigms, one unresolved space

Every prompt buries the chat input further. The interface introduced a conversational model without resolving its conflict with the navigational one already there.

Undisclosed latency kills features

"Thinking..." communicated busy — not progress, duration, or interactability. The design of the wait matters as much as the wait itself.

The fulfillment earned what the entry point didn't

Every time users reached AIM output, the reaction was positive. The product had a strong core. The interface was preventing users from reaching it.

"Users didn't distrust the AI. They distrusted the interface around it."

Abstract

The quality of AI synthesis is not determined by the model alone.

This paper proposes a structured framework for AI-assisted qualitative synthesis that maintains the traceability and rigor of grounded theory methodology while operating at the speed and scale that modern research practice demands.

The framework addresses a gap that current AI tools do not solve: the absence of a verifiable chain of custody from raw participant data to synthesized insight. It was developed from practice between 2022 and 2026.

The central argument is twofold. First, the quality of AI synthesis depends primarily on the structure of the data entering the system. Second, human presence is not optional in AI-assisted research — it is a foundational requirement. Without a researcher who has witnessed the data, there is no basis for validating what the AI produces.

"If you cannot trace it back to the source, it does not belong on the plate." — The mise en place principle, applied to research.

The Problem

AI tools appear trustworthy regardless of input quality.

AI language models are designed — at the interface level — to appear trustworthy. They deliver outputs with confidence and fluency regardless of the quality of the input. When researchers feed unstructured, noisy, or poorly organized data into these systems, the output can appear rigorous while being partially or wholly fabricated.

This is not a failure of the technology. It is a failure of the methodology surrounding it. Most practitioners have responded in one of two ways: accepting AI output without verification, or rejecting AI assistance entirely. Both responses are inadequate.

A third path exists: designing the data structure and human presence that makes AI synthesis verifiable. This is what the framework proposes.

The insight that drives the framework

AI tools constrained to a well-defined input produce more reliable output than AI tools given unconstrained access to raw data. The framework applies constraint through data architecture and structured prompting — making any AI tool more defensible.

Theoretical Grounding

Built from practice. Validated by theory.

This framework was developed independently through practice before its connections to existing methodology were recognized. In January 2026, the author encountered grounded theory methodology (Strauss & Corbin, 1998) for the first time and found that the pipeline built through trial and error over three years was structurally identical to what Strauss and Corbin had formalized decades earlier.

The parallel is significant. Grounded theory was developed precisely because researchers recognized that imposing analytical frameworks on qualitative data before the data had been examined produced findings that reflected the researcher's assumptions rather than participants' actual experiences. The same risk applies to AI-assisted synthesis.

The framework adapts principles from grounded theory for large language model-assisted synthesis — including bottom-up coding from data, researcher immersion, constant comparison, and data saturation.

Core Concept

An atomic observation is one value. One piece of data.

The unit of analysis in this framework is not the sentence. It is the value-bearing claim. A sentence can contain one value or many. The researcher's job — and the AI's job under constraint — is to decompose meaning until each unit is irreducible.

The decision rule: Can this observation be split further without losing a distinct piece of meaning? If yes — split it.

One observation = one value = one evidence quote = one evidence type.

Common observation types that emerge across research contexts: Preference, Emotional state, Behavioral pattern, Causal relationship, Contradiction, Social validation, Expectation. This is not a fixed taxonomy — different research contexts will surface different types.

The backlog principle

When a statement cannot be cleanly decomposed, it is not discarded — it is preserved with "ambiguity" as its evidence type. If the same unresolved signal appears across multiple participants, the pattern itself becomes attributable. The framework is not anti-ambiguity. It is pro-defensibility.

Path A — Generative Research

The data reveals the themes. You don't impose them.

For generative research (interviews, exploratory sessions), the pipeline follows an inductive path. Codes emerge bottom-up from atomic observations. The data generates the code.

Structured Transcript → Clean Notes → Anonymization → Atomic Observations → Evidence Type → Inductive Codes → Axial Codes → Themes

Structured Transcript

Raw transcript + observation notes combined. Speaker labels added, metadata included, observation notes embedded at the exact moments they occurred.

Clean Notes / Reconciliation

AI removes filler, normalizes statements, replaces facilitator questions with neutral context lines. Researcher observation notes stay unchanged.

Anonymization

All names and sensitive content replaced based on pre-defined anonymization notes. Nothing disappears silently — redactions are flagged.

Atomic Observations

Each piece of clean data broken into atomic observations — one idea, one source, no interpretations. Each receives a unique code that travels unchanged through every subsequent stage.

Evidence Type / Taxonomy

Applied after all observations exist — never during extraction. The AI classifies against a taxonomy the researcher designed before the pipeline ran.

Inductive → Axial → Themes

Codes emerge bottom-up. Inductive codes cluster into axial codes — higher-order patterns visible across participants. Axial codes cluster into themes that name what the data has revealed.

Path B — Evaluative Research

The data measures against a known standard.

For evaluative research (user testing), deductive clusters are defined by the research design before the pipeline runs. The UX Area is the deductive code — brought in from outside, not discovered from the data.

Structured Transcript → Clean Notes → Anonymization → Atomic Observations → Evidence Type → UX Area → Cluster Description → Findings → Recommendations

The critical reconciliation step

In user testing, data cleaning is a genuine analytical task — not housekeeping. A participant who says "this is clear" while clicking the wrong button is telling us something important with their behavior that their words are not. The researcher makes that judgment call. The reconciled version becomes the clean note. The AI works from this truth.

What It Produces

A structured, queryable knowledge asset — not a deck.

The output of this pipeline — whether themes or findings — is defensible, filterable, and honest.

✓

Defensible

Every conclusion traces to its source. No finding exists without attributed evidence. The chain from recommendation to participant quote is unbroken.

✓

Filterable

Because observations are structured in a spreadsheet with coded columns, the data can be filtered by participant, session, topic, evidence type, UX area, or business risk.

✓

Honest

Because the pipeline requires human presence at every validation gate, the output reflects what participants actually said and did — not what the model inferred they probably meant.

"Traceability creates auditability, not epistemic validity. What the pipeline guarantees is not that the output is correct — it is that the output is checkable."

Implications

The researcher role is not diminished. It is elevated.

Critical thinking moves upstream — into data architecture decisions — and downstream — into validation against lived experience. The researcher who develops these skills becomes more strategically valuable. The researcher who does not risks producing synthesis that looks rigorous and cannot be defended.

On hallucinations: this pipeline does not eliminate them. No methodology does. What it changes is detectability. Errors that previously passed unnoticed become visible at the validation gates, because the structured chain of custody gives the researcher something concrete to verify.

The pipeline does not prevent AI from making mistakes. It makes mistakes findable. That is the real reliability guarantee.

About the author

Ignacio Cánovas is a UX Researcher and AI Experience Designer at Accenture Song, where he developed and applied this framework across research projects spanning automotive, financial services, global search behavior, and enterprise AI agent development. Before technology, he spent nearly a decade in Michelin-starred kitchens across Europe and the Americas — where the standard was simple: if you cannot trace it back to the source, it does not belong on the plate.

Ignacio
Cánovas

Accenture Reinvention
Navigator

Google Chrome
AI Mode Integration

Two types of searchers. Which one are you?

How observations became defensible findings

Finding the button was just the beginning

Where users started shaped their first impression. Nothing more.

Once they landed on the SRP, the mental model converged

When search behavior collides with AI chat interface

Great shortcuts when they worked. A dealbreaker when slow.

Two entry points. One set of problems at the destination.

Organizational Change
at Scale

Design set up to fail from the start.

AI-accelerated discovery. Human-validated conclusions.

18 interviews. 50+ documents. One through-line.

Five recommendations to reposition design as strategic.

From design as a service to design as a strategic force.

A big shift in how the org thinks about design.

What this means beyond this engagement.

Data Mise en Place

The quality of AI synthesis is not determined by the model alone.

AI tools appear trustworthy regardless of input quality.

Built from practice. Validated by theory.

An atomic observation is one value. One piece of data.

The data reveals the themes. You don't impose them.

The data measures against a known standard.

A structured, queryable knowledge asset — not a deck.

The researcher role is not diminished. It is elevated.

IgnacioCánovas

Accenture ReinventionNavigator

Google ChromeAI Mode Integration

Two types of searchers. Which one are you?

How observations became defensible findings

Finding the button was just the beginning

Where users started shaped their first impression. Nothing more.

Once they landed on the SRP, the mental model converged

When search behavior collides with AI chat interface

Great shortcuts when they worked. A dealbreaker when slow.

Two entry points. One set of problems at the destination.

Organizational Changeat Scale

Design set up to fail from the start.

AI-accelerated discovery. Human-validated conclusions.

18 interviews. 50+ documents. One through-line.

Five recommendations to reposition design as strategic.

From design as a service to design as a strategic force.

A big shift in how the org thinks about design.

What this means beyond this engagement.

Data Mise en Place

The quality of AI synthesis is not determined by the model alone.

AI tools appear trustworthy regardless of input quality.

Built from practice. Validated by theory.

An atomic observation is one value. One piece of data.

The data reveals the themes. You don't impose them.

The data measures against a known standard.

A structured, queryable knowledge asset — not a deck.

The researcher role is not diminished. It is elevated.

Ignacio
Cánovas

Accenture Reinvention
Navigator

Google Chrome
AI Mode Integration

Organizational Change
at Scale