User-First Intelligence

Who created User-First Intelligence, and where has it been used?

Who is behind this, in 30 seconds

Venkata Pagadala

Creator of User-First Intelligence

10+ years in search

AT&T today; before that Apartments.com (CoStar Group) and American Addiction Centers (Rehabs.com).

Applied at enterprise scale

Put to work in enterprise search, including at a Fortune 50 company.

Peer-reviewed research

Springer Nature's Discover Artificial Intelligence and the Journal of Digital & Social Media Marketing.

Presented on stage

Industrial Level Classification with Intent, brightonSEO San Diego 2026.

Book in preparation

User-First Intelligence, the framework on this page.

Not just talk: below is the method running on my own data, the talk itself, and what people in the field said about it.

1. ImplementedLive knowledge graph↓
2. PresentedbrightonSEO San Diego 2026↓
3. Reviewed10 peer quotes↓
4. PublishedTwo peer-reviewed papers↓

Implemented: my LinkedIn, mapped as a knowledge graph

AI and LLMs is what I publish most (75 posts), yet its median post drew 23 reactions, fewer than my median post overall (27). Products and building, SEO and search, AI search (GEO and AEO), Marketing and content, Data and automation and Life and community each drew more per post. Career and learning and Tech and industry news had fewer than ten posts each, too few to read. The decision is not to stop writing about AI; it is to write about AI the way the stronger areas work: through things I built, data I measured and lessons a reader can use. Posts published 19 October 2025 to 19 October 2026, reactions counted on 19 October 2026; a post can sit in more than one area. Descriptive, not causal.

“Supply is not demand, even on one profile.”

Presented: brightonSEO San Diego 2026

I first presented the method on 15 September 2026, in Track 1 at brightonSEO San Diego 2026, under the line "User First, Algorithm Second". This is the recording, with the slides and the room as they were that morning.

▶ Play the full talk here (15 min)Full recording: Industrial Level Classification with Intent, brightonSEO San Diego 2026Close ✕

On stage in Track 1. More photos, the slides and everyone who showed up are on the talk's recap page.

Reviewed: what people in the field said

Before and after the talk, people who work in search wrote about it in public, and others have written about the research behind it. These are their own words, each linked to where it was said.

What is User-First Intelligence?

User-First Intelligence (UFI) is a framework for understanding demand from the person outward. It starts with who people are, follows them to the platforms where they spend time and ask for help, collects what they need in their own words, and places all of it on one map of an industry, so a business can decide what to build, publish or fix for each audience, in the format and on the platform that audience prefers.

The name states the order of work. User first: begin with the person, not the keyword, the ranking or the model. Intelligence: turn scattered signals into decisions, with the evidence attached. The algorithm still matters. It comes second, because algorithms change every few months and the needs behind them change far more slowly.

“Start with the person. The algorithm comes second.”

How is User-First Intelligence different from SEO and keyword research?

A keyword tool sees only the part of demand that reaches a search box in a countable phrasing. People also ask in Facebook groups and on Reddit, watch demonstrations on YouTube and TikTok, read reviews before they buy, and increasingly put the whole question to an AI assistant. Search is one side of the coin.

“I'm not saying traditional tools are bad, but they're only good for direction. The volume is not a strategy.”

UFI does not replace SEO. Search engines remain the largest demand source most businesses can measure, and technical SEO still decides whether a page can be found at all. UFI decides which needs deserve a page, a video, a product feature or nothing, and for whom; SEO, answer engine optimization and generative engine optimization make the chosen answers easy to find.

Why does User-First Intelligence matter now?

1. 1.People ask in more places. A homeowner with a leaking roof might ask an AI assistant, read an AI summary on Google, watch a contractor on YouTube and post in a local Reddit thread before calling anyone. Every new platform that takes off sends a keyword list out of date.
2. 2.AI answers sit between the question and the page. When an AI summary appeared, people clicked a traditional result in 8% of visits, against 15% without one.
3. 3.Volume is a model, not a census. A search volume is a tool's estimate for one phrasing, one market and one period. When I was choosing a $60,000 three-row SUV for my young family, the answers I trusted came from a Facebook group of owners: which trim, which dealer, what they wished they had known. Questions like those rarely register as volume, yet AI engines answer them by reading exactly those communities.
4. 4.Audiences move. Among U.S. teens, Facebook use fell from 71% in 2014-15 to 32% in 2024 (Pew Research Center, 2024). A plan built on last year's platform map is a plan for last year's audience.

“Zero volume does not mean zero demand.”

“I don’t care about volume. I care about, is there one user that is going to buy my product? I'm going to answer them.”

What are the principles of User-First Intelligence?

How does the User-First Intelligence framework work?

{
  "id": "rec_000184",
  "text": "Which trim fits two car seats and adults in the third row?",
  "surface": "community_thread",
  "source_url": "https://example.com/thread/0184",
  "collected_at": "2026-10-23",
  "role": "demand",
  "language": "en",
  "entity": "three_row_suv.seating",
  "intent": "compare",
  "journey_stage": "consider",
  "persona": "parent_young_children",
  "label_confidence": 0.86
}

Where do people ask? The twenty demand sources

I presented the twenty-source map at brightonSEO San Diego 2026 and wrote it up for Previsible. Follow the need across sources: if it appears in queries, community threads, video comments and reviews, those views should inform one decision, not four departmental reports.

▶ Click through the deck from hereSlides: Industrial Level Classification with Intent, brightonSEO San Diego 2026Close ✕

“So these four slides will be the most important slides in this entire presentation.”

Which platform and which format for which persona?

Among adults aged 18 to 29, 63% use TikTok and 58% use Snapchat, against 12% and 4% of those 65 and older. So if your audience is 18 to 25, short video on TikTok and Snapchat may reach it far better than a blog post. Check the prior against your own analytics: the same persona may still finish the purchase after a search.

“And content doesn't mean that textual content. It can be video, audio, podcast, newsletters, whatever it may be.”

How do you build the industry map? Ontologies, knowledge graphs and intent

Ontology: design the classes before the data arrives

An ontology defines what kinds of things exist in an industry and how they relate: products, problems, needs, places, people, providers. Design it top down, from broad category to specific need, with modifiers that cut across every branch. Write the competency questions first, the questions the map must answer, such as "which storm-damage needs, in which towns, have no good answer?" (Ren et al., 2014). See the ontology topic.

Knowledge graph: connect the entities to everything around them

A knowledge graph holds the specific entities (products, brands, places, conditions, competitors) and the links between them. Google brought the idea to search in 2012 as "things, not strings" (Google, 2012). In UFI the graph reaches past the company: an industrial context graph covers the industry and its adjacent markets, so a rival arriving from next door appears on the same map. See the knowledge graphs topic and the graph types guide.

Source roles: keep evidence in its lane

Each source gets exactly one role. Structure says what exists. Coverage says who answers. Answerability says whether the best answer is good enough. Demand says who asks. Coverage never counts as demand: that one rule prevents the most common error in content research, copying what competitors publish and calling it what people want.

Intent: why the person is asking

Every record is classified twice: where it sits on the map, and why the person is asking: to research, compare, purchase, get support or reach something they already know. Broder (2002) separated navigational, informational and transactional web searches; what is new is that the same words carry different intent in a search box, an AI assistant, a forum and a video comment, so UFI labels intent per surface and records named entities as it goes.

@prefix ufi:  <https://venkatapagadala.com/ufi#> .
@prefix owl:  <http://www.w3.org/2002/07/owl#> .
@prefix rdfs: <http://www.w3.org/2000/01/rdf-schema#> .

ufi:Need          a owl:Class .
ufi:StormDamage   a owl:Class ; rdfs:subClassOf ufi:Need .
ufi:HailDamage    a owl:Class ; rdfs:subClassOf ufi:StormDamage .
ufi:Town          a owl:Class .
ufi:Page          a owl:Class .

ufi:about     a owl:ObjectProperty ; rdfs:range ufi:Need .
ufi:inTown    a owl:ObjectProperty ; rdfs:range ufi:Town .
ufi:answers   a owl:ObjectProperty ; rdfs:domain ufi:Page ; rdfs:range ufi:Need .
ufi:role      a owl:ObjectProperty .   # structure | coverage | answerability | demand

PREFIX ufi:  <https://venkatapagadala.com/ufi#>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>

SELECT ?town ?need (COUNT(?ask) AS ?asks)
WHERE {
  ?ask  ufi:role ufi:Demand ; ufi:about ?need ; ufi:inTown ?town .
  ?need rdfs:subClassOf* ufi:HailDamage .
  FILTER NOT EXISTS {
    ?page ufi:role ufi:Coverage ; ufi:answers ?need ; ufi:inTown ?town .
  }
}
GROUP BY ?town ?need
ORDER BY DESC(?asks)

Where does retrieval fit? From embeddings to AI answers

Inside: retrieval sorts demand onto the map

Embeddings turn each question, review and post into a vector, and a vector index finds the nearest nodes of the ontology in milliseconds. The expensive step, a model's decision, runs only on a handful of candidates, with the graph supplying each node's definition, neighbors and examples. This is retrieval-augmented generation turned inward (Lewis et al., 2020): the pattern that grounds a chatbot in documents grounds a classifier in an industry map, and graph-based retrieval extends it to questions about the whole corpus (Edge et al., 2024). Sort into a map you designed rather than clustering: clusters describe the data you happened to collect, not the industry.

Outside: AI engines retrieve before they answer

Google describes its AI Mode as using a query fan-out technique: it breaks a question into subtopics and issues many searches at once on the person's behalf (Google, 2025). A brand is chosen by the sub-questions, not only the head question, and each engine reads different kinds of sources, so map how each need fans out (try the Fan-out Journey) and measure presence answer surface by answer surface.

# Pass 1: retrieval narrows each record to a few candidate nodes
q = embed(record.text)                        # one vector per record
candidates = index.search(q, k=8)             # cosine similarity over node vectors

# Pass 2: a model decides, with the graph's context and a fixed output schema
context = [graph.describe(node) for node in candidates]   # definition, neighbors, examples
label = llm.classify(
    text=record.text,
    candidates=context,
    schema={
        "node": "one of the candidates, or NONE",
        "intent": ["research", "compare", "purchase", "support", "navigate"],
        "confidence": "0 to 1",
        "reason": "one sentence",
    },
)

# Quality: agreement with a gold sample labeled by people
kappa = cohen_kappa(gold.labels, model.labels)   # report with n and an interval

How do you find the gaps that matter?

Gaps are an outcome of query discovery, the framework's first task: find the questions people ask about the industry's entities, on any surface and in their own words, before deciding which to answer. A Content Gap Score turns the comparison into a ranking; its open reference implementation is published on GitHub.

Evidence 1: who AI answers cite

YouTube alone took 12.4% of all citations, twice the leading brand's 6.2%, and third-party platforms together out-cited the 15 leading brands combined. Be present where the engines read: the video, the community answer and the reference entry, not only your own pages. Brand names are masked; the platforms are named because they are the finding.

Evidence 2: what customers praise and what they complain about

The top complaint, rude staff, appears in 4.4% of reviews, and those reviews average 1.91 stars; long waits appear in 2.8% and average 2.39. No blog post fixes a wait. Fix the experience first: digital check-in and in-app appointments, then a page that tells people what to expect and how to book. Reviews of one U.S. telecom retailer's stores; the retailer is not named, and themes were assigned by a classifier, several per review where they apply.

“None of the traditional tools tell a problem or pain points, but analyzing and classifying data tells that.”

prior(e) = β·rel(e) + (1 − β)·dem(e)
CGS(e)   = prior(e) · (1 − own(e)) · [ α·(1 − cov_K(e)) + (1 − α)·mean_q(1 − ans(q,e)) ]

rel(e)    topic-sensitive PageRank over the structure graph, rank-normalized    (structure)
dem(e)    rank-normalized count of user questions linked to e                   (demand)
own(e)    the organization's own coverage, graded 0, ½ or 1                     (coverage)
cov_K(e)  mean of the same grade over the competitor set K                      (coverage)
ans(q,e)  whether the web answers competency question q about e                 (answerability)

How do you see where the audience is moving?

UFI watches the persona and platform links over time, and a short list of early signals:

“It's a moving target. You need to find where your users are spending the most time.”

What happens when every result is personal?

Search results already vary from person to person, and AI answers vary more. A 2013 audit of Google found that personalization changed about 11.7% of results on average, driven mainly by being signed in and by location (Hannak et al., 2013). AI assistants compose their answers instead of ranking pages: ask the same question twice and the answer, and its sources, can change, and researchers at OpenAI found that a chatbot's responses can differ with nothing more than the user's name (Eloundou et al., 2025). As assistants remember more about each user, that variation will grow.

“What your tools see is not what your users see.”

Here is where I think this goes. These are predictions, not findings:

“Agents gather. People decide.”

How do you run a User-First Intelligence study?

1. Step 1

   Frame the decision and write the questions first

   Decide what the study must inform (a content plan, a product, a market entry) and write the competency questions the map must answer before collecting any data.

2. Step 2

   Define the personas and grade the evidence

   Record each persona's situation, tasks, constraints and channel habits, and mark every trait measured, derived or inferred.

3. Step 3

   Map platforms and attention

   For each persona, note where it spends time, where it asks and where it trusts answers. Use survey data as the prior and your own analytics as the check.

4. Step 4

   Collect demand in people's own words

   Pull questions, posts, comments, reviews and queries from the platforms on the attention map, keeping each record's source, role and date. Keep competitors' coverage apart from demand.

5. Step 5

   Classify everything onto one industry map

   Place each record by entity, need, intent, persona and journey stage. Let AI agents do the first pass, check a gold sample by hand and fix the map where it fails.

6. Step 6

   Compare demand, supply and competitors to find the gaps

   For every node, compare the need with your coverage, your competitors' coverage and the quality of the best available answer, surface by surface, and rank the gaps by value and effort.

7. Step 7

   Decide, ship, measure and refresh

   Choose what to build, publish or fix, in which format and on which platform. Measure completed tasks and business outcomes, re-run the collection on a schedule and watch where people move.

“When you're working for an enterprise, you need to build a system at scale that can be used by multiple users.”

Worked example: evidence-graded personas on this site

The personas on this site are built this way, from 3 named studies with published methods, samples and field dates: 6 traits measured and 3 inferred, shown rather than hidden. Behind them, the Global Persona Framework (version of 23 September 2026) lists 463 dimensions in 42 families, and states its own rule: identity does not determine behavior or interests. The Persona Builder turns the framework's questions into a brief, and the Fan-out Journey shows one persona's need becoming dozens of questions across a purchase.

How do you measure User-First Intelligence?

Where did User-First Intelligence come from?

1. August 2024

   User-First Optimization

   I described programmatic SEO as UFO, User-First Optimization, plus rich data, plus scaled implementation: what value do you give the user that is better than your competitors?

2. April 2025

   Users stay constant

   UFO became my stated strategy: algorithms, devices and results pages come and go; users stay constant, whether the engine is a search engine or an answer engine.

3. September 2026

   User first, algorithm second

   At brightonSEO San Diego I presented Industrial Level Classification with Intent: industry-wide demand research connecting queries, including zero-volume ones, to personas, intent and pain points. The slides, the write-up on Previsible and a recap post are public.

4. October 2026

   User-First Intelligence, named and specified

   The framework took its name, and an open reference implementation of its Content Gap Score was published on GitHub.

5. 23 October 2026

   This reference guide, version 1.0

   A research paper and a book on User-First Intelligence are in preparation. Until they are out, this page is the framework's public reference; it is versioned, and changes will be noted here. UFI is my independent work, not a product or a position of my employer.

The recording of that talk, photos from the room and what people in the field said about it are at the top of this page.

How does User-First Intelligence relate to existing ideas?

1. 1986

   User-centered design

   Norman and Draper put the user at the center of system design. UFI applies that stance to demand research and content.

2. 1999

   Personas

   Alan Cooper made users concrete for product teams. UFI grades every persona trait by its evidence.

3. 2002

   A taxonomy of web search

   Broder separated navigational, informational and transactional searches: intent, classified.

4. 2010

   People ask their networks

   Morris, Teevan and Panovich showed people put questions to their social networks, not only to search engines.

5. 2012

   Market structure from text

   Netzer, Feldman, Goldenberg and Fresko mapped a market's structure from forum text.

6. 2014

   Competency questions

   Ren et al. used competency questions to decide what an ontology must answer.

7. 2016

   Jobs to be done; competition from search data

   Christensen, Hall, Dillon and Duncan asked what a customer hires a product to do; Ringel and Skiera mapped competition among more than a thousand products from search data.

8. 2019

   Customer needs from user-generated content

   Timoshenko and Hauser found customer needs in user-generated content with machine learning.

9. 2024

   Generative engine optimization

   Aggarwal et al. studied how sources gain visibility in AI-generated answers; Google's people-first content guidance points the same way.

10. 2026

    User-First Intelligence

    The combination, run as one system: a designed industry map every demand source lands on; evidence with roles and grades; persona, platform and format decided together; agents at scale; and a loop that watches where people move.

Where can you watch, read and try User-First Intelligence?

Conversations about the ideas

Play

Voices of Search (Previsible) · 19 October 2026 · 44 min

Rethinking SEO for LLMs and Personalization

With host Tyson Stockton: how enterprise search is changing with LLMs. Mining first-party data and zero-volume queries, building topic hierarchies, and getting ready for personalized search and query fan-out.

Closest to this guide's sections on zero-volume demand and personalization.

Play

SEO Video Show · 10 October 2026 · 43 min

How to Find What AI Search Trust

Episode 302: the SEO data most teams are missing, and finding demand beyond Google.

Closest to the twenty demand sources: finding demand beyond Google.

Play

AI SEO Show · 21 October 2025 · 33 min

Enterprise AI SEO: Automation Pipelines, MCPs, and Future Proofing

How enterprise teams ship automation pipelines for metadata, schema and content drafts, classify queries at scale, and earn inclusion inside LLM answers.

Closest to the industry map: classifying queries at enterprise scale.

Play

The Free Coalition, Expert Series · 15 October 2026 · 59 min

Building AI Agents for Marketing Functions with Venkata Pagadala

A session for marketers on building AI agents: how agents differ from scripts and workflows, when one agent is enough and when a multi-agent system makes sense, and how to test, measure and guard agents before scaling them.

Closest to the rule that agents gather and classify while people decide.

Each player loads from YouTube only when you press play.

Tools you can use today

The LinkedIn knowledge graph above is one live tool. These four run the persona side of the method, in your browser, on public data.

Persona BuilderAnswer the framework's follow-up questions and grade the evidence for each answer as you go. You get a persona brief with no invented biography.Build a persona →Persona Fanout JourneyOne starting query, fanned out by persona scenario and by every stage of the buyer's journey. Every question says which inputs produced it and how to check it.Fan out a query →Global Persona FrameworkThe classification library behind the personas: 463 dimensions in 42 families, each with its definition, its follow-up question and the evidence a value needs.Browse the framework →Audience PersonasPersonas where every trait is graded measured, derived or inferred, and links to the study behind it. Build anyone and see where they actually are.See the personas →

Research papers

Frequently asked questions about User-First Intelligence

How to cite User-First Intelligence

@misc{pagadala2026ufi,
  author       = {Pagadala, Venkata},
  title        = {User-First Intelligence (UFI): A Framework for Understanding People Before Algorithms},
  year         = {2026},
  month        = oct,
  note         = {Version 1.0. Reference guide},
  howpublished = {\url{https://venkatapagadala.com/guides/user-first-intelligence}}
}

Quote the one-sentence definition with attribution and a link to this page. When the paper and the book are published, their details will be added here.