Vivander Resources · Personal AI guides · Part 3

Will Your Organization Make a Customer’s AI Agent Shortlist?

How people prompting AI and personal agents discover businesses differently, and what to assess before committing your 2027 budget.

See what the audit assesses
Magnifying glass revealing a hotel through glass panels, illustrating how AI examines businesses
Conceptual illustration: discovery starts with clear business information and verifiable evidence.

Imagine a customer asking their personal AI assistant to find a hotel for an anniversary weekend. They want somewhere quiet, a good bar, a particular neighborhood, and a total price within their budget. They ask it to compare three options and prepare the next step for their approval.

Your organization now has several opportunities to enter that decision. It also has several opportunities to fall out of it. The system needs to discover you, understand your offer, check the customer's requirements, and find a usable route to act.

That is why Vivander has updated its AI visibility audits to factor in how personal agents search for businesses on their users' behalf. Alongside the answers people receive when they prompt AI, Vivander examines the evidence and paths to action that support delegated research.

This third installment explains the difference, the research behind it, and how to establish where your organization stands.

The customer supplies a goal. The system can build a workflow.

A person may type a question, read the answer, and decide what to investigate next. In a delegated task, the system can select searches and tools, inspect the results, and choose another step while working toward the person's goal.

OpenAI's July 2025 agent launch documented a system combining research with browser interaction, text browsing, and API access. It also described user confirmation for consequential actions. This is historical evidence of those capabilities, rather than a current product setup guide. [1]

Personalization adds another dimension. Google's January 2026 Personal Intelligence announcement described opt-in use of Gmail and Photos context in AI Mode, including travel plans and preferences. That establishes a personalization mechanism; AI Mode itself should not automatically be classified as a persistent autonomous agent. [2]

For a business, the practical implication is worth testing: the words in one search query may represent only part of the customer's requirements.

How human prompting and agent research differ

The following table describes common workflow differences. Capabilities overlap, and a conversation can move from a direct answer into delegated work.

DimensionPerson prompting AIAgent working for the personBusiness implication to test
Starting pointA question or request, with context the person provides.A goal, constraints, available tools, and any authorized context.Can your offer be matched to a specific need?
QueriesThe person writes the visible prompt and follow-ups. AI may generate searches underneath.The system can divide the goal into searches and tool requests.Can individual facts be found when searched separately?
PreferencesPreferences can appear in the prompt or conversation.Preferences may also come from saved context or connected sources, where supported and permitted.Are the attributes that determine fit explicit and accurate?
Evidence gatheringThe person may inspect citations and websites.The system can retrieve sources, inspect pages, and revisit missing details.Do your organization’s website and public listings agree on important facts?
Next stepThe person often decides which link to open or action to take.The system may attempt a supported action or hand it back for approval.Can it reach the appropriate booking, buying, or inquiry route?
MeasurementAnswer inclusion, citations, factual accuracy, and referrals.Those measures plus constraint satisfaction, how far the task progresses, and where it fails before the booking or inquiry handoff.Does the audit capture what happens after a mention?

What the research actually shows

1. AI-assisted choices can involve fewer website visits

The Nuremberg Institute for Market Decisions reported a study comparing people using ChatGPT with people using Google for shopping-related tasks. Its summary says 12% of ChatGPT participants checked other websites, compared with 73% of Google participants. These are participant-level results from constrained tasks, and no autonomous agents were tested. [3]

Vivander’s interpretation: businesses should examine the information presented before a customer arrives on their website. Website analytics alone provide an incomplete view of that earlier consideration.

2. One request can generate several searches

Google documents that AI Overviews and AI Mode may issue related searches across subtopics and sources while building an answer. This is commonly called query fan-out. [4]

For visibility testing, Vivander would therefore examine the separate questions inside a buying decision. A hotel could be considered on location, then require another check on cancellation terms or parking costs. A restaurant could be discovered for cuisine, then require evidence about private dining capacity.

3. Similar results can come from different search journeys

An April 2026 preprint compared 39 people with one GUI-agent configuration on ten tasks in a production audio-streaming application. The agent's navigation was more centered on search and involved less branching; people explored content more. This small, application-specific study supports examining behavioral differences. It cannot establish how every personal agent will choose a hotel or restaurant. [5]

Vivander’s interpretation: test the route an agent actually takes. A polished homepage may be only one of the places it encounters your organization.

4. Agents can generate instructions, and developers can optimize them

The Automatic Prompt Engineer research showed that a system could propose and evaluate instructions, achieving human-level or better performance on 24 defined instruction-induction tasks. The human baseline consisted of existing written instructions. It was not an equal-budget contest against expert human prompt optimizers. [6]

Three activities need separate labels: generating a search query, delegating a subtask to another model, and systematically testing reusable prompts against a metric. An agent may do the first two during a task. Automated prompt optimization requires an evaluation process. An automatically written instruction is not automatically an optimized instruction.

For business owners, the implication is straightforward: a customer may express a need once while the system produces several working instructions along the way. Publishing variations of one imagined customer prompt leaves much of that journey untested.

A hotel example: follow the decision through

Magnifying glass revealing a hotel through glass panels, illustrating a Vivander AI visibility audit
Conceptual illustration: inspect the evidence behind the hotel shortlist.

This is a fictional illustration, not a client result or a recorded agent trace.

Find three quiet hotels for an anniversary weekend in Charleston. Prioritize a strong bar, a central location, and flexible cancellation. Keep the total two-night cost below $1,000. Verify the terms and show me the options before booking.

A person might ask this question and review the answer. An agent with suitable tools could take the following steps:

  1. Discover candidates. Search for hotels matching the destination and desired experience.
  2. Investigate fit. Look for evidence about location, bar facilities, and the quiet-stay preference. Treat claims about quietness as uncertain when evidence is insufficient.
  3. Check current terms. Inspect date-specific rates, mandatory charges, and the cancellation rules of the actual offer.
  4. Compare. Explain which options meet the requirements and identify facts it could not verify.
  5. Prepare the handoff. Present the appropriate booking route and obtain any required approval.

Possible searches could include “Charleston hotel cocktail bar,” followed by a named hotel's cancellation policy or parking charges. A short query can sit inside a much richer task.

A restaurant version might involve a party size, dietary requirements, a date, seating preferences, and reservation availability. A professional-services version might involve geographic coverage, relevant experience, engagement terms, and an inquiry route.

What Vivander’s updated audits establish

Vivander retains the four visibility states: invisible, mentioned, cited, and recommended. Vivander applies them to the recorded test conditions. A recommendation in one run describes that run; it does not establish universal visibility.

For personal-agent discovery, Vivander also assesses five practical questions:

  • Discovery: Does the business enter consideration for relevant customer tasks?
  • Understanding: Are its offer, location, audience, and distinctive strengths represented correctly?
  • Verification: Can the system locate reliable evidence for decision-critical claims?
  • Customer fit: Does the proposed choice satisfy the defined constraints, and are unknowns disclosed?
  • Action: Can the system reach a usable next step? Vivander records whether it found no route, found a link, reached a form, or reached the final step before submission.

Each test records the platform, the model where available, the mode, the date, the request, the location, any customer profile used, and the observed failures. Every request runs three times. Customer profiles are written for the test and labeled as test conditions.

Standard tests run without saved memory or account history so results can be compared over time. A small, labeled set adds a written customer profile to show how stated preferences change the answer.

The audit stops at the booking or inquiry handoff. It never submits a reservation, a payment, or a lead form.

What to improve first

Vivander’s recommendation is to begin with the facts that determine whether a customer can choose you:

  1. Make the offer specific. Explain what you provide, who it serves, where it is available, and the conditions that apply.
  2. Publish decision-critical details. Identify the relevant prices, fees, capacity, accessibility features, policies, and service limits. Date information that changes.
  3. Resolve conflicting public information. Assign an owner to correct material inconsistencies across your organization’s website, business profiles, and booking or selling platforms.
  4. Check access and readability. Inspect what supported systems can retrieve and what browser-based tools can actually navigate.
  5. Test the handoff. Follow the customer task into the booking engine, checkout, reservation service, or inquiry form.

Google's guidance for its AI search features retains foundational SEO requirements, including crawl access, useful textual content, and structured data consistent with visible information. It specifies no special schema requirement for AI Overviews or AI Mode. Those statements concern Google's features; other systems need their own checks. [4]

Vivander would prioritize observed failures over speculative technical rebuilds. Fix a missing cancellation policy or broken booking route when the evidence identifies it as a problem.

Establish the baseline now. Budget the work for 2027.

Four glass panels connected by a rising line, illustrating AI visibility monitoring over time
Conceptual illustration: compare repeated checks and investigate what changes.

Vivander’s view is that personal agents increase the urgency of understanding how your organization is represented. They expand the workflow Vivander needs to assess, adding research, preference matching, and potential action to the discovery question.

The research reviewed here does not supply a representative, current three-way field comparison of people using traditional search, people prompting AI, and persistent personal agents on identical commercial tasks. It also does not establish a universal conversion advantage. That uncertainty makes a business-specific baseline valuable.

The audit identifies and measures gaps. Your organization’s team owns implementation and maintenance. Use the following distinction when assigning responsibilities and budgeting for 2027.

What the audit measuresWhat your organization’s team maintains
Baseline assessment: Recorded answers, sources, constraint checks, and booking-path tests.Current business information and the website, booking, and inquiry routes assessed.
Content and source accuracy: Missing decision-critical information and conflicting public facts.Content updates, source corrections, and ownership of accurate public information.
Access and handoff: Observed access, readability, navigation, and booking or inquiry failures, with the last step reached recorded.Technical fixes and functioning booking or inquiry paths.
Repeatability: Three runs of every request, with test conditions and observed failures recorded.Ongoing source maintenance and scheduling future assessments as information changes.
Commercial scope: Visibility and progress to the booking or inquiry handoff. The audit does not track or validate completed transactions.Referral tracking, consented customer research, and verification of actual inquiries or transactions, where the client chooses to undertake this work.

A mention, a citation, a recommendation, and a purchase are separate outcomes. A useful budget connects the gaps you observe to the work you can verify.

Research and source notes

  1. OpenAI: Introducing ChatGPT agent, July 17, 2025. Historical provider description of research and action capabilities. The launch page now marks itself as outdated.
  2. Google: Personal Intelligence in AI Mode in Search, January 22, 2026. Provider description of opt-in personalization, rather than independent adoption or conversion evidence.
  3. NIM: Tech, Trust, and Deception, 2025. Originating organization's summary of human shopping-task research; the external-site visitation figures do not measure autonomous agents.
  4. Google Search Central: AI features and your website. Documentation on query fan-out and website eligibility, reviewed October 13, 2026.
  5. Movin et al.: Same Outcomes, Different Journeys, April 2026 preprint. A small study in one production audio-streaming application.
  6. Zhou et al.: Large Language Models Are Human-Level Prompt Engineers, Automatic Prompt Engineer research, revised 2023. Defined instruction-induction tasks and fixed human-written baselines.

My perspective: opportunities and limitations

Potential benefitsCosts and limitations
Clearer business information can support both customer understanding and agent research.Corrections across multiple sources require ongoing ownership and maintenance.
Testing specific customer needs can reveal overlooked content and handoff gaps.Personalization and system changes make results variable.
Booking-path tests can uncover friction beyond the initial recommendation.Tool limitations, access restrictions, and approval steps can interrupt progress.
A baseline can make 2027 spending more targeted and reviewable.Improved answer visibility alone does not establish revenue impact.