Skip to content
Outrigger ResearchMethods pilot · August 2026 · 14 min read

The same AI visibility test named different winners.

In a 1,000-question pilot, changing how we collected ChatGPT answers changed whether businesses appeared visible—and, in some markets, which business appeared to be winning.

By Joel House & Outrigger Research · Exploratory engineering pilot

Analysis ID: decision-divergence-exploratory-v1

The measurement problem

A dashboard can count every answer correctly—and still measure the wrong customer experience.

The question is no longer whether an AI visibility score is mathematically accurate. It is whether the answer-collection method represents the product experience the buyer believes it represents.

What changed

The disagreement was large enough to change the marketing diagnosis.

We compared the API with web search against an automated ChatGPT Search workflow. The strongest result was not a wording difference. It was a change in whether a business appeared visible at all.

16.4%

visibility disagreement

82 of 500 business-market comparisons

27.5%

local disagreement

the largest segment-level gap

14.6%

disjoint winner sets

14 of 96 usable market contexts

32.1%

recommendation overlap

across 800 unnamed questions

“Closer” and “interchangeable” are different standards.

Finding one

The methods disagreed about visibility in one out of six comparisons.

Our panel contained five selected businesses in every market context: a market leader, major challenger, value provider, premium or specialist provider, and smaller provider. For each business, we looked across eight unnamed buying questions and asked a simple question: was this business recommended at least once?

That produced 500 business-by-market comparisons. The API with web search and ChatGPT Search gave different answers in 82 of them: 16.4%, with a 95% confidence interval of 12.8% to 20.2%.

Classification resultComparisonsRate
Both methods agreed41883.6%
One recommended the business; the other did not8216.4%

We made the test stricter by requiring at least two appearances among the eight questions. The disagreement rate barely moved: 16.6%. The result was not simply created by isolated one-off mentions.

Finding two

Sometimes the apparent market winner changed too.

We counted how often each of the five sampled businesses was recommended across the eight questions in its market. In 96 contexts, both methods recommended at least one panel business. Their winning sets had no business in common in 14 of them, or 14.6%.

Some contexts had ties. When we isolated the 48 contexts where both methods produced one clear winner, they named different businesses in 11 cases, or 22.9%.

That number needs a sensitivity warning. When we required each winning business to appear on at least two of eight questions, the estimate fell to 16.3%. At a four-of-eight threshold, it fell to 6.9%. The lesson is not that 22.9% is a universal error rate. It is that the winner changed often enough to matter—and platforms should show how much evidence supports a ranking.

Three examples

What a changed winner looks like to a marketer.

01

A New York pest-control company went from invisible to the panel winner.

The API with web search recommended Control Exterminating on four of eight questions and OnGuard on none. ChatGPT Search recommended OnGuard on three and Control on two.

API with webAPI diagnosis: OnGuard has a fundamental discoverability problem and should benchmark Control.

ChatGPT SearchSearch diagnosis: OnGuard leads the sampled customer experience and should learn how to defend it.

A business cannot sensibly be treated as both invisible and the market leader without explaining the measurement route.
02

The email-marketing leader changed even though both brands were widely visible.

For a growing 20-person business, the API recommended Brevo on six of eight questions and ActiveCampaign on five. ChatGPT Search recommended ActiveCampaign on all eight and Brevo on seven.

API with webAPI diagnosis: use Brevo as the primary competitive benchmark.

ChatGPT SearchSearch diagnosis: ActiveCampaign holds the strongest sampled position.

Both dashboards can contain correctly counted data while creating different competitor priorities.
03

The trust leader changed for a high-consideration mortgage purchase.

The web-enabled API recommended Better Mortgage on three privacy-and-trust questions and Rocket Mortgage on one. ChatGPT Search recommended Rocket on four and Better on three.

API with webAPI diagnosis: Better owns the strongest association with privacy and trust.

ChatGPT SearchSearch diagnosis: Rocket owns the strongest sampled association.

That difference can redirect messaging research, digital PR, comparison pages and sales battle cards.

These examples do not prove which company a consumer would ultimately choose. They show that the two collection methods would not produce the same market diagnosis or action list.

Finding three

Local businesses had the largest measurement problem.

Local services27.5% · 160 comparisons
Consumer products12.5% · 120 comparisons
High-consideration12.5% · 80 comparisons
B2B software9.3% · 140 comparisons

Local answers can be sensitive to the city, precise location, review sources, directories, opening hours and whichever pages are retrieved at that moment. The two methods also received location context differently. This pilot cannot identify one cause, but it shows why a method validated on software brands should not automatically be assumed reliable for dentists, plumbers or moving companies.

Web search improved the API—but did not make it interchangeable.

Without web search, the API disagreed with ChatGPT Search about sampled-business visibility in 21.8% of comparisons. With web search, that fell to 16.4%. When both methods produced one unambiguous panel winner, the exploratory different-winner estimate fell from 41.7% to 22.9%.

This is not an argument against APIs. A scalable API can be appropriate for frequent directional monitoring. A smaller customer-facing panel can then calibrate whether that scalable method still represents the experience it is intended to measure.

Share the research

Five findings, ready to quote.

Share the finding directly, copy the accompanying soundbite, or download the full-resolution 1200×675 card for X, LinkedIn, newsletters and presentations.

AI visibility’s hidden measurement problem

outriggerai

1 in 6

visibility classifications changed

The API with web search and ChatGPT Search disagreed in 82 of 500 sampled business-market comparisons.

1,000-question pilot

The gap was largest for local businesses

outriggerai

27.5%

visibility disagreement in local services

More than one in four local business comparisons changed from visible to absent—or the reverse.

1,000-question pilot

Same questions. Different winner.

outriggerai

22.9%

named a different clear panel winner

In the 48 contexts where both methods produced one clear winner, they selected different businesses 11 times. At a stricter threshold, the estimate was 16.3%.

1,000-question pilot

Web search helped. It did not close the gap.

outriggerai

32.1%

average recommendation overlap

Across 800 unnamed buying questions, the two methods returned exactly the same recommendation set only 5.1% of the time.

1,000-question pilot

The standard AI visibility tools now need

outriggerai

Measure

the customer experience you claim to measure

A score needs a measurement contract: collection route, web access, evidence threshold, uncertainty and calibration.

1,000-question pilot

A practical buying standard

Five questions every buyer should ask an AI visibility vendor.

01

Which experience are you actually collecting?

The answer should name the model or customer-facing product, whether web search is used and whether collection happens through an API or rendered interface.

02

How do you validate the scalable method?

If most questions run through an API, ask how often the results are compared with the customer-facing product.

03

Can I inspect the original evidence?

A buyer should be able to examine the question, answer, recommendation list, sources, collection time and collection route behind a score.

04

How much evidence is required before declaring a winner?

Ask about repeated prompts, ties, minimum thresholds and uncertainty. One appearance should not silently become a decisive ranking.

05

Does reliability vary by market?

A single platform-wide accuracy number can hide weak performance in local search or another sensitive category.

What Outrigger measures

Your visibility score should come with evidence, not just a percentage.

Outrigger shows the prompts, answers, sources and competitors behind the score—and uses customer-experience calibration to keep scalable monitoring honest.

See your real baseline

Find out what AI says about your brand—and why.

Run a free AI visibility audit

Research evidence and citation notes

The citable record.

This was an exploratory engineering pilot, not a preregistered causal experiment. The complete claim boundaries are included so the findings can be cited without overstating what the data establish.

Study population

  • 1,000 unique buying questions
  • 800 unnamed questions in the main comparison
  • 25 business categories
  • 100 category-context cells
  • Five sampled businesses per context
  • 3,000 successful answers and extractions

Collection routes

  • API without web search
  • API with web search required
  • Automated ChatGPT Search workflow
  • Same frozen prompt corpus
  • Surface-blinded recommendation extraction
  • Business registry normalization
MeasureResult95% confidence interval
Visibility disagreement82/500 (16.4%)12.8%–20.2%
Repeated-visibility disagreement83/500 (16.6%)13.0%–20.6%
Disjoint winner sets14/96 (14.6%)8.3%–21.9%
Different clear winner11/48 (22.9%)12.5%–35.4%

Limitations that matter

  1. The decision-divergence measures were designed after the original results were known.
  2. Each question was run once per method, so route effects remain mixed with ordinary model and retrieval variation.
  3. The automated Search workflow still needs validation against visible logged-in ChatGPT sessions.
  4. The recommendation extractor was blinded to route, but independent human coding remains outstanding.
  5. The winner comparison covers a frozen five-business research panel, not every provider in each market.
  6. The study measured descriptive decision divergence—not actual marketing actions or financial harm.

Download

Aggregate results (.csv)

Includes primary estimates, sensitivity thresholds and segment results.

Analysis ID

decision-divergence-exploratory-v1

Suggested citation

House, J., & Outrigger Research. (2026). The Same AI Visibility Test Named Different Winners: An Exploratory Decision-Divergence Analysis of a 1,000-Question API-versus-ChatGPT Search Pilot. Outrigger AI. https://www.outriggerai.com/research/chatgpt-api-vs-search

Prompt corpus SHA-256: a22a37c2592d8d3714aaffdb6ccb301ba9bbdbcd5d82dd3a50f3fe79209ec535

ChatGPT API vs. Search: A 1,000-Question Study