Find
Retrieve candidate pages, entities, records, services, locations and sources relevant to the goal.
The simplest useful model
Find → Establish → Verify → Choose → Act
An AI agent is not simply a chatbot with a browser. The important change is the feedback loop. The system can decide what it needs next, use a tool or search route, observe what comes back, update its working representation and continue.
Establish
Resolve which real-world thing is being discussed and connect major relationships around it.
Verify
Test claims against evidence, provenance, source reliability, external records and contradictions.
Choose
Compare candidates and determine which representation best satisfies the user’s intent.
Act
Use tools to contact, calculate, book, request, organise or hand the next step to a person.
Important: this five-stage model is a human-readable abstraction. It is not a claim that every proprietary AI product uses the same hidden internal architecture.
From response generation to investigation
The loop is the agent.
Foundational research such as ReAct showed how reasoning and actions can be interleaved. Toolformer explored how language models can decide when and how to call external tools. Reflexion demonstrated how feedback can be retained and used to change later attempts.
Modern research systems push that pattern much further. OpenAI describes Deep Research as planning and executing multi-step trajectories, backtracking and reacting to information it discovers. Google’s DeepSearchQA benchmark explicitly tests fragmented-source collation, entity resolution and stopping decisions. Microsoft’s AgenticRAG equips a reasoning model with search, find, open and summarise tools so that retrieval becomes iterative rather than fixed.
A practical agent loop
A real implementation may be much more complex, but the observable pattern is recognisable:
Plan
Search / Act
Observe
Update
Repeat
At any point the system may branch, backtrack, ask for more information, call a different tool or stop when it believes the task is sufficiently resolved.
Why browsing alone is not enough
Retrieval is becoming a journey, not a lookup.
OpenAI’s BrowseComp benchmark is unusually revealing because simply adding browsing to GPT-4o produced only a modest improvement. The purpose-built Deep Research system performed dramatically better on the same difficult, multi-hop tasks.
0.6%GPT-4o accuracy on BrowseComp
1.9%GPT-4o with browsing
9.9%OpenAI o1 without browsing
51.5%Deep Research on BrowseComp
OpenAI notes that the Deep Research model was specifically trained for persistent browsing tasks. The comparison therefore illustrates the value of strategic reasoning and tool use rather than a neutral head-to-head product ranking.
| Research | What it tests | Why it matters here |
|---|---|---|
| BrowseComp | Hard-to-find, entangled open-web information | Shows that browser access alone is not enough. Search-path selection and interpretation matter. |
| DeepSearchQA | 900 tasks across 17 fields | Tests fragmented-source assembly, entity de-duplication and when an agent should stop searching across 900 prompts and 17 fields, while documenting premature stopping and low-confidence over-retrieval. |
| AgenticRAG | Iterative retrieval over knowledge bases | Microsoft reports the biggest ablation gain came from moving from single-shot retrieval to agentic tool use. |
The DMS solar-system model
“What is this?” 🌞🪐🌙
One way to visualise an agent establishing an unfamiliar entity is as a solar system. The Sun is the core entity or proposition. Planets are substantial supporting data nodes. Moons are narrower contextual details. Evidence supplies the gravitational field that determines whether the emerging representation holds together.
The system does not have to discover the Sun first. It might enter through a small contextual fact, resolve the larger subject it belongs to, then move outward again in search of corroboration.
Solar system ≠ literal model internals. It is a visual explanation of the information problem: central entities, major relationships, contextual details and the evidence needed to stabilise them.
Sun = core entity or proposition · Planets = substantial supporting nodes · Moons = contextual support · Gravity = evidence, provenance, reliability and consistency.
Evidence gravity
Information alone does not hold the system together.
A website can publish thousands of claims. That does not make the resulting entity model trustworthy. The important question is what happens when an investigative system challenges what it has ingested.
Large system, weak gravity
A publisher generates huge quantities of highly structured assertions:
- “Best specialist in Devon”
- “Award-winning expert”
- “Verified by multiple websites”
- Thousands of repeated machine-readable relationships
If the important claims cannot be independently recovered, the representation may be large but poorly grounded.
Smaller system, stronger gravity
The core relationships point toward evidence that survives inspection:
- Business identity resolves consistently
- Credential can be checked with the issuer
- Location is supported by independent records
- Projects, services and claims have recoverable evidence
The graph is not necessarily the evidence. It can be the index to the evidence.
ProvenanceWhere did the claim come from, who produced it and what was it derived from?
Source independenceAre apparently corroborating sources genuinely distinct, or copies of one publisher?
RecoverabilityCan the important relationship be independently rediscovered outside the original graph?
ConsistencyDo names, services, locations, dates and relationships broadly agree across sources?
ContradictionsWhat happens when an authoritative source disagrees with the emerging representation?
FreshnessIs the relationship still current, and can the system distinguish old evidence from current reality?
Provenance needs layers
Who built the graph is not the same question as whether the entities are real.
A graph may have one identifiable architect or publisher without every entity represented within it sharing that provenance. For open-web investigation, those distinctions should be explicit rather than blurred together.
Publisher provenanceWho constructed, maintained or published the dataset, page cluster or relationship graph?
Entity provenanceDoes the represented organisation, person, service or place independently exist outside the publishing architecture?
Relationship provenanceWho asserts that a particular relationship exists between two entities or claims?
Evidence provenanceWhat independently controlled evidence supports, qualifies or contradicts that relationship?
Common architecture is not the same as common reality. A network can have identifiable publishing provenance while still representing independently existing organisations, addresses, credentials, people, projects and records. The experimental question is whether significant relationships remain recoverable beyond that architecture.
Negative grounding
Failure to recover expected corroboration does not prove a claim false. However, an unsuccessful verification attempt can leave a proposition unresolved. A capable investigative system may then search further, qualify the claim, reduce reliance on it or refuse to use it in a consequential decision.
That is different from contradiction. Contradiction introduces conflicting evidence. Negative grounding describes the informational effect of attempted corroboration that does not succeed.
The motorway model
Solar systems organise information. Motorways connect them.
A traditional website tree implies one preferred route from the homepage downward. Investigative retrieval can be much less obedient. An agent may join through any useful node, move across entities and topics, leave for external evidence, then re-enter elsewhere.
Roads into the ecosystemSearch results, location pages, technical articles, structured data, directories, project evidence and machine-readable resources can all become entry points.
Roads out matter tooA useful information environment should make it easier to reach genuine external records, councils, professional bodies, corporate information and other independently controlled evidence.
The city of information hypothesis
What happens when useful roads become an environment?
The solar-system model describes a task-relevant representation around an entity. The motorway model describes traversal between entities, topics and evidence. The next step is to zoom out again and ask whether repeated, useful relationships can form a broader information environment.
🏠
Pages
Individual documents, records, service pages and evidence resources.
🏢
Entities
Businesses, places, credentials, services, people and organisations resolved from those documents.
🛣️
Relationships
Meaningful roads between entities, claims, services, locations and external sources.
🏙️
Information environment
A larger ecosystem in which useful onward routes repeatedly appear.
🌍
External evidence
Independent records and sources that can confirm, qualify or contradict what the environment declares.
If an investigative agent repeatedly encounters a coherent information environment that resolves entities, exposes useful relationships and leads outward to recoverable evidence, does it become computationally advantageous to continue investigating within that environment?
This is not a claim that an AI system “trusts” a network, nor that a particular graph receives ranking preference. It is a testable question about information value, investigative friction and whether some environments yield more useful evidence per retrieval action than others.
The zoom-out matters: document → entity → neighbourhood → ecosystem → regional information environment. The experiment is not whether a network can repeat claims loudly enough. It is whether a structured environment can make accurate investigation easier while also making important claims easier to challenge.
Where the metaphors meet real research
Graphs, memory, information gain and trust are all active research problems.
No single paper proves the DMS solar-system or motorway models. The striking part is that many of the components our models require already exist independently across modern retrieval, graph, memory and provenance research.
Microsoft GraphRAG
GraphRAG extracts entities and relationships from documents, creates higher-level communities and can enter through relevant entities before expanding into connected graph information and source text.
Microsoft GraphRAG researchDRIFT Search
DRIFT combines community-level information with local search and generates follow-up questions that progressively move toward more specific information.
Microsoft DRIFT SearchAgenticRAG
Microsoft’s 2026 work lets a reasoning model repeatedly search, find, open, navigate and summarise information rather than accepting a single fixed retrieval set. In its reported ablation study, moving from single-shot retrieval to agentic tool use was the largest factor, with a 5.9× improvement.
AgenticRAG paperInfoGain-RAG
EMNLP 2025 research measures a document’s contribution according to how much it changes the model’s confidence in producing the correct answer, rather than similarity alone. The paper reports a 15.3% average improvement on GPT-4o across its evaluated datasets.
InfoGain-RAGReliability-Aware RAG
Another EMNLP 2025 system explicitly combines relevance with estimated source reliability because a relevant source can still be wrong.
Reliability-Aware RAGMAGMA and GAM
ACL 2026 work represents agent memory using relational graph structures, including semantic, temporal, causal, entity and topic relationships, then retrieves through graph-guided traversal.
MAGMA · GAMGoogle Knowledge Vault
Google’s 2014 Knowledge Vault combined web extraction with existing knowledge and probabilistic inference to compute calibrated probabilities of fact correctness.
Google ResearchKnowledge-Based Trust
Google researchers later explored estimating source trustworthiness from factual correctness rather than relying only on popularity-style signals, applying the method to 2.8 billion extracted facts across 119 million webpages.
Knowledge-Based TrustW3C PROV
The W3C formalised provenance around the entities, activities and agents involved in producing information, noting its use in assessing quality, reliability and trustworthiness.
W3C PROV overviewSelected published results: DeepSearchQA contains 900 prompts across 17 fields and highlights persistent problems including premature stopping and low-confidence over-retrieval. Microsoft reports a 5.9× improvement from the shift to agentic tool use in its AgenticRAG ablation. InfoGain-RAG reports a 15.3% average improvement on GPT-4o across its evaluated datasets. These results belong to their respective experimental settings and should not be generalised into universal public-web performance claims.
The restraint matters: Microsoft GraphRAG usually operates over a defined corpus. AgenticRAG often assumes a defined knowledge base. Memory graphs are typically internal. Deep-research agents roam the open web. Similar mechanisms do not prove that every public AI product constructs the same graph over every website.
The deeper investigation model
An investigative agent can construct a provisional representation, not merely collect pages.
The most useful synthesis of current agent, retrieval, graph and provenance research is that an investigative agent can progressively construct a task-relevant representation of the world, then use actions, retrieval and feedback to test and refine that representation until it is sufficient to choose what to do next.
The Sun need not come first
An agent may first encounter a narrow “moon”: a foundation condition, project record, credential or local fact. It can work upward to the larger relationship, resolve the central entity, then travel outward again for corroboration.
Relationships can weaken
The emerging model is not sacred. If later evidence conflicts with an earlier relationship, a capable system can revise the hypothesis, reduce confidence, branch into another search or discard the relationship.
Not every road deserves travel
Agents operate with finite context, latency, tool calls and compute. Useful investigation therefore includes deciding which unresolved relationship is worth inspecting next and which evidence is unlikely to change the answer.
Conceptual investigation loop
User goal→Initial hypothesis→Find entities→Discover relationships→Retrieve observations→Evaluate evidence→Revise / search again / choose / act
This is a conceptual synthesis, not a claim about proprietary hidden chain-of-thought or a universal internal architecture.
The economics of investigation
Useful information environments may reduce investigative friction.
A capable agent does not have infinite retrieval budget. Tool calls take time. Documents consume context. Branches create more branches. Noisy evidence can increase work without improving the answer.
This creates a useful research question: does a coherent information environment repeatedly yield high-value entities, relationships and onward evidence routes per retrieval action? InfoGain-RAG provides a research analogue by measuring whether a retrieved document actually improves generation confidence and correctness. It does not prove that public agents literally calculate a universal “information gain per click” score.
Working hypothesis: a well-connected information environment may become useful to an investigative agent when it reduces the work needed to construct and test a sufficiently accurate representation. More pages are not automatically better. The relevant variable may be useful evidence gained relative to search, context and noise cost.
Two very different 500-page ecosystems
Environment A: hundreds of repeated claims, weak provenance, few external routes and substantial duplication.
Environment B: explicit entities, stable identifiers, meaningful relationships, source provenance and roads to independently controlled evidence.
Both can be large. Only experimentation can establish whether one materially reduces agent investigation cost in a given system.
Declared edges versus recoverable edges
A graph becomes more interesting when the agent can leave it.
Suppose a graph declares that Company A holds Credential B. Repeating that edge across twenty related sites does not create twenty independent votes. The stronger test is whether the agent can travel beyond the graph and independently recover the relationship from the issuing organisation or another genuinely independent source.
Declared edge
Graph → Company A → Credential B
The graph tells the machine the relationship exists. This may aid discovery and navigation, but the declaration alone does not establish independent corroboration.
Externally recoverable edge
Graph → Company A → Credential B → leave graph → issuer record → relationship recovered
The graph functions as an index to evidence. The machine is given enough structure to challenge the assertion rather than being asked to trust it.
Recoverability principle: the strongest machine-readable environment does not need to say “trust this graph”. It can effectively say: do not trust it blindly; here is enough information to go and check it.
Proposed experimental evaluation framework
Turn the concepts into observable tests, not invented scores.
The concepts in this page suggest properties that can be tested experimentally. DMS is not claiming that Google, OpenAI, Microsoft, Anthropic or any other proprietary system internally calculates these dimensions, and we are not presenting them as a validated composite score.
The stronger approach is to define a task, control the information environment where possible, record the agent and tools used, and measure an observable outcome. That allows the framework to be challenged, repeated and improved rather than hidden behind an arbitrary number.
| Experimental property | Possible observable measure | What it would test |
|---|---|---|
| Entity resolution | Correct entity identification rate across controlled trials. | Whether the agent distinguishes the intended real-world entity from similarly named or nearby alternatives. |
| Discovery | Successful discovery rate from different realistic entry points. | Whether the information can be found when the agent starts from a service, place, credential, project or other partial clue. |
| Traversability | Successful completion rate for predefined one-hop, two-hop and three-hop relationship tasks. | Whether meaningful relationships can be followed without repeated dead ends. |
| Recoverability | Percentage of declared relationships independently reconstructed after the originating source is excluded. | Whether a relationship survives outside the graph or publisher that first declared it. |
| Provenance | Correct identification of claim origin, issuer or production history. | Whether the agent can establish where important information actually came from. |
| Source independence | Correct classification of genuinely independent sources versus mirrors, syndication or common provenance. | Whether repeated publication is mistakenly treated as independent corroboration. |
| Evidence gain | Change in task correctness, calibrated confidence or unresolved uncertainty after a retrieval step. | Whether the newly retrieved information materially improves the investigation rather than merely adding volume. |
| Contradiction handling | Percentage of controlled contradictions detected, investigated and correctly resolved or qualified. | Whether the agent revises a neat but wrong representation when evidence disagrees. |
| Geographic resolution | Correct entity-place-service relationship rate at predefined geographic levels. | Whether the agent resolves town, district, county and service-area relationships accurately. |
| Freshness | Correct classification of current, superseded and stale information. | Whether old evidence is distinguished from present reality. |
| Noise cost | Irrelevant or redundant documents, tokens or retrievals processed before the task is resolved. | How much information has to be discarded before useful evidence emerges. |
| Search cost | Median tool calls, searches, opens, elapsed time or retrieval steps required to reach a predefined completion threshold. | How much investigative work is required before the agent can act with sufficient grounding. |
No aggregate “agent score” yet. Combining these heterogeneous properties into one number would require a defensible weighting model, repeatable benchmarks and validation. Until that exists, reporting the underlying observed measures is more rigorous than manufacturing a single score.
| Conceptual layer | Relevant experimental properties | Core question |
|---|---|---|
| Representation 🌞 | Entity resolution, geographic resolution | Has the system constructed the right task-relevant representation? |
| Investigation 🛣️ | Discovery, traversability, search cost | Can it reach the information without wasting excessive investigative effort? |
| Evidence gravity 🧲 | Recoverability, provenance, source independence, evidence gain | Does the representation survive inspection beyond the original declaration? |
| Critical revision 📋 | Contradiction handling, freshness, noise cost | Can the system notice when its current explanation should be revised? |
What the research does not let us claim
Agents are capable investigators, not omniscient ones.
Current systems still stop too early, retrieve noisy evidence, misjudge source credibility, confuse entities and remain vulnerable to hostile web content. Better architecture reduces these problems. It does not erase them.
Entity errors
Wrong initial entity linking can send later graph traversal down the wrong road.
Source judgement
Web access and citations do not guarantee perfect credibility assessment.
Stopping mistakes
DeepSearchQA documents both premature stopping and over-inclusive low-confidence answers.
Prompt injection
Web agents can encounter malicious instructions embedded in external content, so data should be data, not attempts to command the agent.
What this means for a business
AI visibility may only be the entrance exam.
A business can be visible to AI and still be difficult for an investigative agent to establish, verify or use. Agent readiness therefore extends beyond “mention my company in ChatGPT”.
🧭
Can the agent resolve the business?
Names, locations, domains, organisation details and other identifiers should make it easy to distinguish the real entity from similarly named businesses.
🧩
Can it understand the services?
Services, product options, service areas, limitations and supporting pages should form a coherent representation rather than scattered ambiguous prose.
📋
Can it verify important claims?
Credentials, professional relationships, project history and other significant propositions need appropriate evidence routes.
⚙️
Can it actually do anything?
Pricing rules, booking pathways, quote requirements, availability and human hand-offs become increasingly important once the AI is acting for the customer.
David Marketing Specialist already separates these ideas into AI-Agent Ready Business Data, AI Agent Ready Websites, entity resolution, Artificial Geography Optimisation and topical relevance and graphing.
DMS working prototype
From “AI visible” to “AI operable”.
On 27 August 2026, David Marketing Specialist published a machine-readable conservatory service and quotation record for The Plymouth Home Improvement Company. The resource contained price ranges, customer inputs, specification adjustments, feasibility boundaries and explicit rules telling an AI what it could calculate and what required PHIC assessment.
≤5%
Difference from the professional pricing benchmark in the two documented Grok and Google Gemini prototype tests.
This is an observed prototype result, not a claim that all AI systems will retrieve or calculate identically.
Why that experiment matters here
The prototype was not just a schema block saying “PHIC installs conservatories”. It exposed a usable business model:
- published price families and fixed adjustments;
- variable costs that the machine was forbidden to invent;
- foundation and access conditions that require assessment;
- a clear distinction between an indicative machine estimate and a final human quotation;
- supporting location and service resources.
It is an example of the final stage in our model: once an agent can find and understand the business, can it safely perform useful work with the information?
One possible open-web test environment
Apply the model to controlled experiments.
The investigation model is broader than any one DMS project. One practical way to test parts of it is to compare information environments that differ in structure, provenance, external evidence routes and relationship density, then observe how capable agents investigate them.
Example questions worth testing
Can an agent recover the correct entity when it enters through a small contextual clue rather than the homepage?
Does an explicit relationship reduce the number of searches needed to find independently controlled evidence?
When several pages repeat the same underlying claim, does the agent distinguish common provenance from genuine corroboration?
What happens when a harmless contradiction forces the agent to choose between a neat graph and stronger external evidence?
Graphzilla is one test instrument, not the theory
DMS has a separate experimental open-web relationship project internally nicknamed Graphzilla. It may provide a useful environment for some of these tests because relationships, entry points and external verification routes can be deliberately observed.
Its presence here should not be read as evidence that Graphzilla improves AI ranking, retrieval or recommendation. The broader research question is how agents investigate structured and unstructured information environments. Graphzilla is simply one possible place to run controlled experiments.
Experimental discipline: record the model, date and time, prompt, available tools, starting information, retrieval path, sources used, stopping point and outcome. Where relevant, also record location, account or personalisation state, and whether the investigation began from a clean, new or existing session, because these conditions can alter the information available to the system. A failed hypothesis is still a useful result.
Research discipline
Four things we will keep separate.
| Layer | What belongs here | Example |
|---|---|---|
| Research establishes | Published mechanisms and measured results. | AgenticRAG can outperform single-shot retrieval in its tested enterprise settings. |
| DMS observed | Dated retrieval or prototype outcomes from our own tests. | Grok and Gemini interpreted the PHIC machine-readable quotation prototype in the documented scenarios. |
| We infer | An explanation that fits research and observations but is not directly proven. | A coherent entity ecosystem may reduce investigative friction. |
| We test | Questions turned into controlled, repeatable experiments. | Whether a change in information structure measurably alters discovery, recovery, evidence use or investigative cost. |
Selected research sources
Read the underlying work.
These sources support individual mechanisms discussed above. They do not collectively prove the DMS conceptual model, nor do they establish how every proprietary public AI system investigates the open web.
- OpenAI - BrowseComp
Benchmark for difficult web-browsing agents. - OpenAI - Introducing Deep Research
Multi-step web research, backtracking and source synthesis. - Google - DeepSearchQA
Fragmented-source collation, entity resolution and stopping criteria. - Microsoft - AgenticRAG
Iterative search, open, find, summarise and evidence navigation. - Microsoft - GraphRAG
Entity and relationship graph retrieval over defined corpora. - Microsoft - DRIFT Search
Community information plus iterative follow-up questions. - Yao et al. - ReAct
Interleaving reasoning and environment-facing actions. - Schick et al. - Toolformer
Learning when and how to call external tools. - Shinn et al. - Reflexion
Using feedback and episodic reflection across attempts. - InfoGain-RAG - EMNLP 2025
Ranking retrieval by contribution to correct answer generation. - Reliability-Aware RAG - EMNLP 2025
Combining relevance with estimated source reliability. - MAGMA - ACL 2026
Semantic, temporal, causal and entity graph memory. - GAM - ACL 2026
Hierarchical graph-based memory and graph-guided retrieval. - Google - Knowledge Vault
Web-scale probabilistic knowledge fusion. - Google researchers - Knowledge-Based Trust
Estimating source trust from factual correctness. - W3C - PROV Overview
Formal provenance model for web information. - Programmatic Graph Reasoning - EMNLP 2025
Step-by-step fact verification against graph evidence. - OpenAI - Deep Research System Card
Capabilities, web risks and prompt-injection mitigations.
Frequently asked questions
AI agents, without the fog machine.
What is an AI agent?
An AI agent is more than a language model producing one response. In agentic systems, a model can operate inside a loop that plans, uses tools, observes results, updates its working state and decides what to do next.
Do AI agents simply search the web once and answer?
Not necessarily. Modern research agents can search iteratively, inspect sources, reformulate queries, follow entities and relationships, gather more evidence and decide when enough information has been collected.
Does structured data guarantee an AI recommendation?
No. Structured data can reduce ambiguity and expose relationships, but it does not guarantee retrieval, trust, ranking or recommendation. Important claims still need appropriate evidence and may be independently checked.
What does AI-agent ready mean for a business website?
AI-agent readiness means making accurate business identity, services, locations, rules, evidence and customer processes easier for capable AI systems to discover and interpret, while retaining clear safeguards and human hand-offs where required.
Does DMS have a validated AI agent score?
No. The evaluation framework on this page proposes properties that can be operationalised in controlled tests. It does not claim that proprietary AI systems use those measures internally, and it does not combine them into an invented composite score.