Original research · completed August 24, 2026

AI Search Visibility Audit for B2B SaaS: 194-Call Method and Findings

Quoted First's AI search visibility audit for B2B SaaS used 194 analyzable native-search calls across 139 unique query strings, 15 immediate identical repeats, 10 delayed identical repeats, 40 unrelated-category controls, 32 native page opens, and three final 10-query stability batches. Quoted First appeared in zero result blocks; that null is reported rather than reframed as visibility.

194 search calls 139 unique query strings 4,923 retained result blocks No placement claim

Review the exact method Read the findings

Research question

What causes a page to enter a native-search result pool?

The study compared broad and narrow commercial queries, brand and entity wording, exact phrases, natural questions, fresh and evergreen subjects, local services, high-stakes information, niche technical topics, obscure hobbies, and one-variable perturbations.

The goal was behavioral: record what appeared and how stable it was. The study did not attempt to name the private provider, index, crawler, retriever, reranker, embeddings, personalization, or refresh mechanism.

Final accounting

The study exceeded every predeclared minimum.

194

Analyzable native-search calls, one exact query per call.

Longest response requested

4,923

Result blocks in the machine-retained portion.

2,783 unique URLs; first 25 calls excluded

Quoted First result blocks in the research corpus.

Null retained as observed

Research component	Completed	Why it matters
Analyzable search calls	194	One additional travel call had truncated output and was conservatively excluded.
Unique query strings	139	Natural buyer questions, category variants, controls, and one-variable perturbations.
Natural Quoted First/category baselines	15	Supported by extensive buyer-intent, entity, exact-phrase, and domain diagnostics.
Unrelated controls	40	Five calls each across eight categories to avoid fitting the model only to marketing pages.
Immediate exact repeats	15	Tests short-window result-set nondeterminism.
Valid delayed exact repeats	10	Repeated 15–18 minutes later while other research continued.
Native opens	32	Twenty-eight yielded useful page text; four returned a cache miss, timeout, or internal error.
Final stability batches	3 × 10	None materially changed the leading behavioral model.

Download the captured URL ledger (CSV): inspect 3,495 displayed result rows from calls C006–C110, covering 105 calls and 85 unique query strings. This is a machine-readable subset, not the complete 194-call corpus. The 4,923-block retained portion used for final accounting excludes the earliest 25 analyzable calls; neither number is presented as a complete row-level export. Download the captured URL-level subset.

Exact methodology

Declare, search, preserve order, inspect, repeat.

Every placement search used one query per native-search call and requested the longest available response. Displayed block order—not internal reference numbering—defined rank.

Start with real commercial probes

The first five searches covered broad service, managed startup, pre-revenue B2B SaaS, and ChatGPT-citation intent before any framework was built.

Build a natural query ladder

Queries added audience, stage, budget, location, constraints, service versus software, brand ambiguity, exact phrases, and intended outcome without manufactured tokens.

Preserve every ordered result set

Titles and URLs were recorded in the order displayed. Nulls, noisy results, duplicate parameter URLs, and failed later opens were retained.

Change one query variable

Pairs compared best/recommended, question/keyword, broad/narrow, year/no year, location, budget, audience, service/software, singular/plural, and format.

Repeat exact strings and audit equality

Immediate next-call repeats and later repeats were compared by full visible sequence and top-ten URL overlap. Only character-for-character identical queries counted as repeats.

Open only discovered pages

Representative leaders and anomalies were opened through the native tool to inspect title, H1, opening answer, page purpose, audience, specificity, evidence, dates, internal context, URL, and format.

Query-family coverage

Commercial findings were checked against unrelated jobs.

Family	Variables tested	Representative formats
Commercial and buyer intent	Best, service, software, managed, stage, budget, audience, outcome	Homepages, focused landing pages, pricing pages, comparisons
Brand and exact phrase	Exact brand, ambiguous brand, category attachment, misspelling, quoted phrase	Brand sites, dictionaries, PDFs, literal phrase matches
Local and travel	Geography, specialization, room type, neighborhood, constraints	Direct providers, directories, hotel pages, aggregators
Health, legal, and financial	Evidence, official forms, risk, jurisdiction, current guidance	Government pages, statutes, PubMed, professional guidance, PDFs
Technical and emerging	New protocols, official docs, current drafts, implementation details	GitHub, documentation, standards, research papers
Consumer and obscure hobby	Budget, installation constraint, question wording, forum intent	Small specialist publishers, reviews, forums, manuals

Repeatability

Identical queries were usually stable; changed queries were not.

Short-window stability was high enough that a changed word must never be mistaken for time volatility. It was not perfect determinism.

Immediate identical repeats

9 of 15 reproduced the full visible sequence.

Five others kept the same leading core while changing or reordering the tail. One local query changed its number-one result immediately.

Interpretation: direct evidence of high short-window stability with rare nondeterminism—not proof of a deterministic system.

Delayed identical repeats

3 of 10 reproduced the full visible sequence.

Across valid repeats 15–18 minutes later, mean top-ten URL membership overlap was 8.4/10. Every query retained a recognizable semantic core.

Interpretation: delayed ordering was neither deterministic nor random; broad and ambiguous pools refreshed more.

Final stability check: the same 10-query battery ran three times. Eight of 10 complete sequences were identical from batch 1 to 2 and again from batch 2 to 3; the other comparisons retained 9/10 top URLs, and every top-three result stayed fixed.

Representative ordered result set

A narrow B2B SaaS query promoted an exact-purpose deep page.

Exact query: LLM search optimization for a pre-revenue B2B SaaS. The stage modifier was not literal in every leader, but the category and audience job stayed coherent.

Position	Displayed page	Page-purpose observation
1	Clear Cited — AI Search Optimization for B2B SaaS	Focused B2B SaaS service page.
2	Semrush — SaaS AI Search Optimization: 8-Step Playbook	Detailed SaaS workflow article.
3	Tracsio — AI Search Optimization for B2B SaaS Founders	Founder and low-authority stage guide.
4	Clare Videre — Which AI Should B2B SaaS Optimize For?	B2B SaaS decision guide.
5	Thomas Peham — B2B SaaS AI Search Playbook	Professional playbook.
6	DiscoveredLabs — Generative Engine Optimization guide	Broad but detailed optimization guide.
7	DiscoveredLabs — Agentic browsing readiness	Technical B2B SaaS guide.
8	Exalt Growth — 12-metric LLM visibility framework	Framework and template.
9	Mostash — Future of B2B SaaS marketing	Industry/audience article.
10	Leadanic — AEO for B2B guide	Focused B2B guide.

Observed—not causal proof: a focused page from a smaller publisher led a famous domain for this narrow job. That pattern recurred across pricing, managed-service, local-service, hotel-room, official-form, documentation, and forum queries.

Selection findings

The behavioral model separates observations from explanations.

Finding	Evidence label	Boundary
Query constraints changed the returned page-purpose pool. Pricing, official form, forum, family-room, audit, managed, and local-problem wording repeatedly promoted matching deep pages.	Observed	Some budget and stage constraints were dropped when exact supply was scarce.
Title, H1, opening answer, and audience alignment were common among opened leaders.	Observed pattern; high-confidence inference as an eligibility input	Exact phrases could amplify the wrong meaning and were not sufficient alone.
Authority depended on the job. Official sources led health, legal, financial, standards, and technical queries; small specialists led narrow commercial and hobby queries.	Observed; intent-conditioned authority is a strong inference	No universal authority metric was exposed or measured.
Format followed the task. PDFs/forms, GitHub/docs, forums, directories, landing pages, and research databases all led appropriate queries.	Observed	No format won universally.
Result pools were usually stable under identical wording and could jump under one-word perturbations.	Observed	Compatible with pool selection plus ordering, but does not prove a two-stage architecture.
Entity clarity appeared relevant to Quoted First's absence. Five brand variants resolved to dictionaries, literal phrases, or generic category pages.	Observed absence; entity/candidate problem is a strong inference	Crawl status, index status, provider, and exact cause remain unknown.

Null results

Quoted First appeared in zero research result blocks.

The null covers 15 natural category and buyer baselines, extensive entity and domain diagnostics, applicable repeats, and the wider 194-call corpus.

The two-common-word brand was often interpreted literally: dictionaries, quotation guidance, phrase-bearing PDFs, programming questions, and semantically adjacent AI-search pages appeared instead.

What a B2B SaaS audit should do

Measure eligibility and stability before assigning a visibility score.

01 · Declare

Freeze natural buyer questions

Include category, use case, audience, buying stage, constraints, competitors, and success states before the first run.

02 · Separate

Keep evidence states distinct

Absent, mentioned, recommended, cited, and directly linked are different outcomes. A page open is a functional check, not search placement.

03 · Repeat

Confirm exact wording

Repeat apparent hits immediately and later, compare the actual strings, preserve order and URLs, and never explain a changed-query pool as volatility.

04 · Inspect

Study the pages that win

Compare purpose, audience, opening answer, evidence, price, format, limitations, entity clarity, and internal context.

05 · Change

Run one useful intervention

Make the smallest change that answers a real buyer need and remains valuable even if external placement does not move.

06 · Report

Publish contradictions and nulls

Label observations, strong inferences, weak hypotheses, and unknowns. Do not claim access to private ranking internals.

Limitations

What this study cannot establish

The search provider, index, crawler, embeddings, personalization, retrieval stack, ranking features, or refresh schedule were not exposed.
Result behavior cannot prove that there is literally a candidate-pool stage followed by a reranker; that remains a weak architectural hypothesis.
Authority, page purpose, internal links, topical depth, freshness, and entity clarity are correlated on live pages and were not all isolated causally.
The retained block and unique-URL totals exclude the earliest 25 analyzable calls, so they are lower bounds rather than complete-corpus totals.
Four native opens were limited by a cache miss, timeout, or internal error. Their original result blocks still count as observed placement.
The corpus is substantial but finite and time-bounded to August 24, 2026. Search outputs can change after the capture window.
Quoted First publishes and benefits commercially from the study. No customer outcome, independent validation, ranking promise, or causal placement claim is made.

Apply the method—or have the first intervention managed.

Use the public worksheet to build your own evidence record, or evaluate the fixed-scope pilot by its deliverables, price, and limits.

Run the DIY audit Evaluate managed service