What Did Our 97-Record AI Website Audit Show?

By Ryan Goering, founder and CEO of BaaDigi, a contractor marketing agency. Last updated: September 2026.

Correction, September 2026: The original August snapshot has been recovered and its stored counts checked. Earlier conclusions about causation, universal agent usability, or improvement from BaaDigi work remain withdrawn. This is a baseline record review, not a new live scan.

What does the recovered August snapshot contain?

The original client_ai_audit table contains 97 records dated August 13, 2026. We checked the following aggregate fields on September 22, 2026 without publishing client identities.

The AI-named counts and site-status counts are separate partitions of the 97 records. The six undetermined AI observations are not the same group as the six unscannable sites. One record marked unscannable has a positive AI-named flag; one marked clean has undetermined AI status.

For the AI-named field, 70/91 is about 77% when the six undetermined observations are excluded. Across all 97 records, the same 70 is about 72%. Neither denominator should be hidden. Site checks also vary between two and three applicable scored checks, so “clean” is not a claim that every record passed an identical full audit.

What can the snapshot not establish?

This is a convenience sample of client records, not a representative sample of businesses. The surviving records do not include the exact original prompt set required for a controlled repeat comparison. The field name ai_cited must not be interpreted as proof that the business's own website was linked.

The table contains 12 clean-site records marked not named and 30 needs-fix records marked named. Those counts show that the stored outcomes can coexist. They do not prove statistical independence, identify a ranking cause, or establish that off-site work would fix every missing mention. Nor are baseline counts evidence of improvement caused by BaaDigi.

What does the larger audit set show now?

The 97-record snapshot was a one-day convenience sample. Our free audit tool has since completed 1,047 audits for self-selected prospects (BaaDigi client data, Mar 16 to Sep 22, 2026), and the technical picture is similar: 93.7% of sites had HTTPS, 82.0% were mobile responsive, 72.9% had an XML sitemap and 59.5% had any schema markup. The median mobile PageSpeed score was 61 and 22.1% of sites scored under 50. Of the 254 audits that also ran a Perplexity check, the business was named for its own service-and-city query in 59 cases, 23.2%. Those are two separate measures again, and again we do not claim one causes the other. The LLM SEO guide covers what does move the AI side.

How are website usability and AI visibility different?

Usability asks whether a visitor or agent can understand and complete an action on the website. Visibility asks whether a search or answer system shows the business or cites a source. A site might be mentioned in an answer while its inquiry form is difficult to use. A working form does not establish that an assistant will mention the business.

Test those questions separately. A technical check should identify an observed condition and a reproduction step. A visibility check should retain the prompt, engine, date, context and answer. Combining them requires a justified match between the exact business, website and market. Similar names or an assumed city are not enough.

What belongs in a reproducible audit record?

Preserve raw evidence privately when it contains customer or account information. A public article can explain its method and limitations without exposing client identities. Permission to scan a public page does not automatically supply permission to publish a client’s private business records.

What does an automated accessibility check tell you?

It can identify conditions that deserve investigation, such as controls without clear accessible labels. The W3C form-labeling guidance explains how labels connect controls with their purpose. Use that guidance to improve the actual form, then test the task again. Do not replace a missing label with a claim about a guaranteed conversion increase.

An automated result is not a complete accessibility audit. Nor does it establish how every browser agent behaves. Agents can use different combinations of visual information and page structure. The meaningful finding is specific: which control, what test, what happened, and whether the intended action could be completed under those conditions.

How should you handle a domain that no longer represents the business?

Confirm that the returned page belongs to the intended business before treating a successful HTTP response as a successful scan. A domain can load a parking page, a replacement owner or a sales listing. Redirects can also take the test to a different property. Record the final URL and inspect the page identity.

If the intended website is unavailable, retain that fact as a separate outcome. Do not score a substitute page and count it as the business’s site. When a scan times out, record a timeout rather than guessing whether the site is down. These distinctions change the denominator and the meaning of every reported percentage.

Can a technical defect explain a missing AI mention?

Not on its own. Two observations occurring together do not establish that one caused the other. A business may be absent from a sampled answer for many reasons that the audit cannot isolate. Likewise, a technically clean page does not prove that a missing recommendation is caused by off-site information.

Google’s AI-feature guidance does not promise inclusion when technical requirements are met. Use a missing mention as a reason to investigate the question, business information and cited sources, not as evidence for a predetermined rebuild. Ask a supplier to identify which finding their proposed work actually addresses.

What should a useful test of your own site include?

Choose a realistic task: find the service area, understand an offer, contact the business or request an estimate. Specify where the test starts and what counts as completion. Use a controlled test environment or an agreed procedure so you do not create false customer bookings or send unintended messages.

Record visible barriers such as moving controls, unclear choices, misleading labels or an error that gives no recovery route. Test a person’s experience as well as the selected automation. The website platform guides can help with a buying decision, while Next.js website services address implementation. Neither is a substitute for a reproducible finding.

When does a paid review make sense?

A paid review is useful when it resolves a defined uncertainty and produces evidence your team can act on. Ask for the tested tasks, limits, screenshots or records, and the expected handoff. Keep search visibility diagnosis separate from a claim of complete accessibility or security certification. Those are different review scopes.

BaaDigi’s AI Visibility Audit is published at $497. It concerns visibility diagnosis and an action plan, not proof that every AI browser can operate your site. For future research, retain the exact prompt set, run conditions, source answers, and matched website identity before comparing snapshots. The recovered baseline cannot supply details that were never preserved.