Contrarian Takes on AI Search: 6 SEO Hypotheses to Test
A practical framework for turning contrarian AI-search claims into measured tests of citations, retrieval, technical quality, trust and brand representation.
· 14 min read
A 50-prompt test can reveal a more useful AI-search insight than a debate about OpenAI, Gemini or AI valuations: a brand may be mentioned in 18 answers yet receive only two first-party citations. These contrarian takes on AI search give SEOs a concrete way to test why that gap exists and decide whether technical repairs, stronger evidence or better brand representation are the right response.
The source context for this discussion is a Semrush Villa session in Cannes featuring practitioners discussing AI visibility. It should not be treated as proof for every claim below. Instead, this article separates observed answer-engine outcomes from proposed hypotheses that a site owner, agency or SEO team can falsify.
Contrarian does not mean automatically disagreeing with the consensus. A credible contrarian view makes a specific prediction, identifies the evidence that would support it and states what result would weaken it. That distinction matters in AI search, where confident narratives often outrun available attribution data.
What makes an AI-search claim worth testing?
A useful claim has three parts: a mechanism, a measurable outcome and a defined failure condition. “AI prefers authentic content” is too vague to guide work. “A page with dated, first-party implementation data will gain more citations than a rewritten category explainer across 30 comparison prompts” is a testable proposition.
For example, a B2B software company could select 40 prompts across four intent groups: 10 category questions, 10 comparisons, 10 implementation questions and 10 troubleshooting questions. It can then record the result in Google AI features where available, ChatGPT, Gemini and Perplexity on the same day. Answer outputs vary, so the record should include the prompt wording, locale, account state where relevant, answer text, cited domains and cited URLs.
A minimum test record includes:
- Prompt and intent: such as “best endpoint security for a 500-person company.”
- Engine and date: name the product, interface and observation date.
- Visibility outcome: brand mention, first-party citation, competitor citation or no relevant result.
- Accuracy outcome: whether the answer states the category, ideal customer and limitations correctly.
- Site condition: whether the likely supporting page is crawlable, indexable, fast and accessible.
This follows the principle behind Audra’s guide to AI search visibility measurement: a score is useful for establishing a baseline, but it does not by itself demonstrate a causal business result.
Contrarian take #1: Distribution can matter more than model benchmarks
AI benchmarks measure particular capabilities under particular test conditions. They do not directly establish which answer experience will shape a user’s search journey. Google has an established search interface, an extensive web index and high-intent search behavior; OpenAI, Gemini and other providers have different product surfaces, integrations and defaults. These distribution factors can matter even if model-quality comparisons change.
This is not a claim that Google will win every use case, or that benchmark quality is irrelevant. It is a narrower SEO hypothesis: a site that improves its visibility in a large, established search ecosystem may see a different outcome from the same site in a standalone AI assistant.
Test the distribution hypothesis
Choose 30 prompts for which a business already has relevant, published pages. Record whether each engine names the brand, cites its domain and links to a relevant URL. Then improve a limited group of pages—for example, repair broken internal links, clarify page titles and add missing comparison information—without changing the prompt set.
Retest after a pre-set period, such as six to eight weeks. If first-party citations rise in one answer surface but not another, that supports the limited conclusion that retrieval and distribution conditions differ. It does not prove one vendor has a better model, and it does not prove the site change alone caused every movement.
The practical lesson is to avoid treating “AI search” as one channel. Reporting should name the engine, prompt cluster and answer type rather than combining every answer into a single visibility number.
Contrarian take #2: Conventional SEO remains a retrieval prerequisite
The claim that SEO has been replaced by answer-engine optimization is too broad. Google’s documentation says there are no additional technical requirements to appear in AI features; pages must be indexed in Google Search and eligible to show a snippet. Its guidance also says established SEO fundamentals remain relevant to appearing in Google Search, including its AI experiences.
That statement applies to Google’s own AI features, not necessarily to every answer engine. It still creates a clear operational priority: a page blocked from crawling, marked noindex, broken by rendering failures or isolated from internal links is an avoidable weak point before any AI-specific tactic is considered.
Audit the retrieval foundation first
Before commissioning 50 new pages, assess the existing pages that should support priority answers. A five-part audit can cover:
- Crawlability: robots rules, server responses, WAF behavior and JavaScript delivery.
- Indexability:
noindex, canonical selection, redirects and HTTP status codes. - Discoverability: internal links, HTML navigation and XML sitemap consistency.
- Performance: whether meaningful content loads reliably without excessive scripts or blocked resources.
- Accessibility and semantics: useful headings, descriptive links, labelled controls and understandable tables.
A useful hypothetical example: in an audit of 84 client sites, an analyst might find 31 sites with one or more noindex pages listed in their XML sitemaps. That is an illustrative scenario, not Audra research or a customer-study result. Its purpose is to show why sitemap and indexability checks should be reconciled rather than reviewed in isolation.
A local audit tool such as Audra can combine crawl, SEO, performance, accessibility and link checks into one page-level review. The hypothesis being tested is not that every technical fix will earn an AI citation. It is that removing material retrieval barriers gives already-useful pages a fairer chance to be found, selected and represented accurately.
Contrarian take #3: Distinctive evidence may matter more than more AI copy
Generative tools can produce large volumes of competent summaries at low cost. That makes undifferentiated explainers easier to replicate, but it does not prove that every original-data page will be cited more often. The testable version of the claim is that evidence an answer engine cannot easily obtain elsewhere may create a more defensible reason to select a source.
For a SaaS business, that evidence could include anonymized implementation timelines from a clearly defined cohort, documented support patterns or a transparent product benchmark. An ecommerce brand might publish fit measurements, return-reason analysis, material testing or original product photography. The important qualifier is scope: a survey of 17 customers is not a market-wide fact.
Google’s guidance on generative AI content emphasizes producing helpful, reliable, people-first material rather than generating pages primarily to manipulate rankings. It does not guarantee citation treatment for proprietary information. The relationship between original evidence and citations must therefore be treated as a hypothesis, not an established causal rule.
Make evidence safe to repeat
An evidence-led page should identify its source, sample size, date range, collection method, exclusions and limitations. It should also state the narrow conclusion a reader can reasonably draw.
For example, a hypothetical agency could write: “Across 84 websites reviewed between July and September 2026, 31 had at least one XML-sitemap URL carrying a noindex directive.” That statement is clearer and more auditable than “many sites have indexing issues,” provided the methodology explains how pages and directives were counted.
Publish one evidence-led asset and one conventional explainer on closely related topics. Monitor first-party citations, correct mentions, referral visits and assisted conversions for eight weeks. If the evidence page earns no extra selection, possible explanations include weak prompt fit, limited authority, unclear methodology or an engine that does not expose source citations—not proof that original research has no value.
Contrarian take #4: Brand accuracy can matter more than raw mention volume
A brand can be visible and still be poorly represented. An answer may list a company as a generic tool, recommend it to the wrong buyer or repeat an outdated product claim. Raw mentions alone cannot distinguish helpful awareness from inaccurate awareness.
This matters particularly for comparison and recommendation prompts. A cybersecurity vendor described as suitable for “small personal projects” when it serves regulated enterprises has a representation problem even if it appears often. The useful unit of measurement is therefore not only presence, but the accuracy and evidential support of the description.
Score representation across engines
Choose five stable brand facts: category, ideal customer, core differentiator, flagship product and an important limitation. Run 20 prompts across at least three answer engines and assign each answer a simple score:
- 0: absent or materially inaccurate.
- 1: partial or ambiguous representation.
- 2: accurate brand representation without a first-party citation.
- 3: accurate representation supported by a relevant first-party citation.
Useful prompt examples include “What is [brand] best for?”, “Compare [brand] with [competitor]” and “Who should not use [brand]?” The results can identify whether the best action is updating an official product page, repairing inconsistent documentation, publishing clearer limitations or correcting a third-party source.
A brand discovery audit for search, AI and social offers a related way to evaluate whether core facts remain consistent across the wider source ecosystem. The hypothesis is not that consistency forces an answer engine to repeat a message; it is that contradictory or stale source material makes accurate representation harder.
Contrarian take #5: AI agents could increase the cost of trust
The optimistic view of AI agents is that they reduce research and purchasing friction. The counterview is a possible trust recession: if recommendations, reviews and outreach become easier to generate at scale, users may become more cautious about provenance, recency and accountability.
This is a scenario, not an established market outcome. It is likely to vary by decision type. A low-cost consumer purchase may need little verification, while a health, legal, finance, cybersecurity or enterprise-software decision can involve greater downside if an answer is wrong.
Measure verification behavior, not sentiment alone
Tag landing pages that receive identifiable AI referrals where analytics makes this possible, while acknowledging that referral data can be incomplete. Compare those sessions with organic-search sessions over the same 30-day period. Review conversion or qualified-lead rate, return visits, visits to pricing or documentation, and searches for terms such as “security,” “methodology,” “reviews” or “support.”
The trust-recession hypothesis gains support if answer-driven visitors repeatedly seek validation material before converting. It weakens if they convert at comparable rates without additional verification. Either result should be interpreted carefully: audience mix, device, campaign activity and attribution gaps can affect behavior.
The operational response is not superficial trust badges. It is retrievable proof: dated documentation, clear authorship when relevant, methodology notes, change logs, accessible tables, contact routes and explicit limitations. Human judgment retains value because a person can assess whether a generalized answer applies to a particular context.
Contrarian take #6: Better measurement may mean accepting less certainty
AI-answer attribution is incomplete. Engines do not consistently disclose every retrieval event, source-selection factor, prompt volume or user path. A disciplined measurement model should state these limits rather than imply that a weekly citation change proves revenue impact.
That does not justify doing nothing. It means running a bounded operating test with a fixed prompt universe, a change log and a small number of leading indicators. The aim is to distinguish observed movement from unsupported causal stories.
A 90-day operating scorecard
For 50 priority prompts, report baseline and final observations after 90 days. This is a proposed operating model, not a Semrush-panel finding or a universal industry standard.
| Metric | Definition |
|---|---|
| Brand mention rate | Prompts naming the brand divided by 50 |
| First-party citation rate | Prompts citing the company domain divided by 50 |
| Accurate representation rate | Answers correctly stating selected brand facts divided by 50 |
| Retrieval readiness | Priority pages passing agreed crawl, index, performance and accessibility checks |
| Citation gap | Prompts citing competitors but not the brand |
The work log should list pages repaired, evidence published, documentation updated and inconsistencies identified. This makes it possible to compare optimization activity with observed outcomes without overstating certainty. The distinction between establishing a baseline and changing it is explored in AI search optimization vs. visibility audits.
Source mapping should follow answer results, not assumptions
It is tempting to assume that every category needs Reddit, LinkedIn, YouTube or a particular publisher. That assumption is often too broad. A troubleshooting prompt may surface documentation and community posts, while a product-comparison prompt may surface editorial reviews, category pages, video demonstrations or first-party evidence.
The useful method is observation first. For each prompt cluster, record the source types and domains actually visible in the answer results, then identify the brand’s legitimate evidence gap. This avoids treating a channel tactic as a universal solution.
| Prompt cluster | Observed source types | Example gap | Testable response |
|---|---|---|---|
| Product setup | Documentation, videos, communities | No walkthrough for a common setup task | Publish a step-by-step support page and video transcript |
| Comparisons | Reviews, editorial lists, category pages | Outdated positioning | Improve proof pages and seek correction of factual inaccuracies |
| Troubleshooting | Docs, forums, code repositories | Verified answer is trapped in support tickets | Turn the verified fix into crawlable documentation |
YouTube can be a useful format for demonstrations, setup instructions and visual comparisons, but it should not be added merely because video is popular. A 12-minute walkthrough with a clear title, named presenter, transcript and supporting page is easier to evaluate than an unlabelled video. Whether it improves answer visibility remains a test outcome, not a guarantee.
A compact audit for AI-search assumptions
Start with 25 to 50 prompts tied to revenue, reputation or support demand. Pair each prompt with the page that should honestly support the answer, then review five conditions:
- Cited: Is the brand named or a first-party URL cited where appropriate?
- Retrievable: Can search systems access the supporting page without status, robots or rendering barriers?
- Fast: Does the substantive answer load without avoidable performance problems?
- Accessible: Can users navigate headings, tables, links and controls with assistive technology?
- Accurate: Does the answer describe the current product, audience and limitations correctly?
The output should be a prioritized backlog rather than a vanity dashboard. If the answer is inaccurate, strengthen the best source of truth first. If a brand is absent, inspect prompt intent, source patterns, site quality and the evidence available before assuming that more AI-generated copy is the solution.
FAQ
What are the most credible contrarian takes on AI search?
The most credible claims are falsifiable. Examples include distribution mattering more than model benchmarks, technical SEO remaining a retrieval prerequisite, distinctive evidence outperforming commodity copy, and accurate brand representation mattering more than mention volume. Each should be tested with fixed prompts, named engines, citation records and documented site changes rather than accepted as a prediction.
Does Google need the best AI model to win search?
Not necessarily. A search product can benefit from distribution, existing user behavior, a web index and established retrieval systems in addition to model quality. For SEOs, the practical test is whether relevant pages are indexed, eligible and selected in Google’s AI experiences. Benchmark leadership alone does not answer that site-level question.
Will AI agents create a trust recession?
They could increase verification behavior in high-stakes categories, but this is a hypothesis rather than a settled fact. Teams can compare AI-referred and organic-search visitors for documentation use, repeat visits, validation searches and conversion paths. Clear provenance, dates, limitations and human support can help users assess whether an answer applies to their situation.
Is the AI boom creating a bubble or contracting existing markets?
That is mainly an investment question and cannot be resolved through an SEO audit. For search teams, the more immediate question is whether simple informational queries produce fewer website visits while citations, branded searches or assisted conversions change. Measure those outcomes by prompt cluster rather than making unsupported forecasts about AI valuations.
How should SEOs test contrarian assumptions about AI-generated answers?
Define 25 to 50 important prompts, capture baseline answers in named engines and score mentions, citations, accuracy and sources. Make one bounded change, such as repairing indexability barriers or publishing transparent first-party evidence. Retest after a defined interval, maintain a change log and treat correlations as leads for further investigation rather than proof of causation.