AI visible today.
Again tomorrow?

The short answer

Repeat the same buyer questions before treating one answer as a website problem. Start with 3 questions, one AI platform and 7 days. Save complete answers and record mentions, recommendations and citations separately. Inspect a repeated gap, repair an established defect, and repeat the same check.

The practical output is a small evidence register that a marketing leader and web owner can review together. Seven days is a starting window, not a statistical threshold or a promise of stable visibility.

1. Problem: a changing answer can trigger the wrong work

Illustrative situation: on Monday an AI answer links to your service page. On Tuesday it mentions the company without a link. A report labels that a visibility failure and the team rewrites the page. On Wednesday the link returns. None of those observations alone proves a website defect or a successful repair.

Business consequence: the team can spend time on the wrong change and report an improvement it cannot substantiate. This is a hypothetical example, not a measured customer result.

What the research supports

Ahrefs followed roughly 37,500 initial citations across five AI tools for 20 days. In that cohort, 50–83% were absent the next day; some platform differences reflected answers with no citations at all. The study also observed citations returning. These are sample-specific results, not a probability for your company.

Read the original Ahrefs study, published 29 October 2026. It supports caution about single checks. It does not establish that any particular page edit caused a citation change.

What this lens checks

Lens 73 tests whether a recommendation fits a customer. This Lens checks how to record repeated observations before deciding what to repair. It covers one chosen platform and a documented set of buyer questions; it does not estimate market reach, leads or revenue.

Mention

The answer names your brand.

Recommendation

The answer proposes your offer as a suitable choice for the stated need.

Target-site citation

The answer links to your own website as a source. A third-party page naming you is a different observation.

Missing check

The run failed or no answer capture is available. This is not evidence of absence.

2. Verify: run a small, repeatable check

1. Select 3 real buying questions. Use questions the intended customer could ask before choosing a supplier. Record the exact wording, buyer market and language. Keep brand-name fact checks separate from category discovery questions.
2. Freeze the available conditions. Choose one AI product, available model, web-search setting and account configuration. Use a separate fresh conversation each time; record memory or personalisation you cannot disable. Fresh conversations do not remove all personalisation.
3. Make one planned run per question per day for 7 days. Save the first completed answer at approximately the same time. Do not keep rerunning until the brand appears. Record retries and failures explicitly.
4. Save evidence before classifying. Keep the full answer, actual citation URLs, date and time with time zone, product/settings and a stable capture reference. Mark mention, recommendation, target-site citation and whether the answer cites any source.

Expected output: 21 scheduled records for 3 questions. Completed answers and missed checks remain distinguishable. If a model or search setting changes, flag the break and compare those runs separately.

One buying-question template

Replace the bracketed fields with a genuine customer need. Use the resulting exact question unchanged for that series. Create two additional questions from different real buying needs and give each its own ID.

Which companies provide [service] for [customer type] in [market]? Compare suitable options against [two important requirements]. Use current public information and cite the sources supporting your comparison.

This is an editorial starting template, not a neutral sample of every buyer. Keep your company name out of discovery questions if the purpose is to observe unprompted inclusion. Do not attach internal company documents.

3. Assess: count observations with the right denominator

For each question and platform, report answers citing the target website / successfully captured comparable answers, plus completed / planned checks. Example only: 3 cited answers out of 6 captured answers, with 1 of 7 planned checks missing. Do not silently present that as 3 out of 7.

Also report how many completed answers cited anyone. A seven-day observation count describes this test, not a universal probability of being recommended.

4. Fix: repair the demonstrated defect

1. Correct the report first. If failed checks were counted as absence, or mentions were mixed with citations, repair the record and calculation before requesting a website change.
2. Investigate a repeated gap. Open the actual cited sources and the intended landing page. Ask the web owner to check intended public access. Ask the commercial owner whether the page answers the buying question accurately.
3. Repair the established issue. Remove an accidental access restriction, correct an outdated or conflicting claim, or add the missing factual answer beside the relevant offer. Change structured data only when it disagrees with visible facts. Keep a before/after capture and assign an owner.
4. Keep an unresolved result honest. If the page is accessible and accurate, record “no website defect established”. Continue observation rather than inventing a fix or guaranteeing inclusion.

The register identifies evidence to investigate; it cannot tell you why a model chose a source.

5. Retest: verify the repair and observe the answer separately

First reopen the live page and confirm the documented defect is resolved. Then repeat the same questions under the same available settings for a further observation window. Record the deployment time, changed passages and any platform changes.

Compare equivalent questions, settings and completed-check coverage. A returning citation is an observed outcome; it is not proof that your change caused it. Longer observation or a stronger evaluation design may be needed.

Download the seven-day evidence register

Download the seven-day evidence register (CSV)

The UTF-8 register contains 21 planned records. Use captured, failed or missed for run_status, and yes or no for the four observation fields. Record dates as ISO timestamps with a time zone. Replace every bracketed field; store sensitive captures in an access-controlled location.

The register is blank evidence scaffolding, not an account of tests already performed.

Optional Terminal check: record completeness

Requires Python 3. Save the CSV and download the register checker in the same folder. Open Terminal there and run:

python3 check-register.py lens-74-seven-day-check.csv

An untouched template should report 0 captured answers and 21 incomplete planned records. The checker flags missing evidence and contradictory labels, then summarises only fully completed comparable records. It does not validate source truth or judge whether a recommendation is appropriate.

Optional PowerShell check

With Python installed, open PowerShell in the folder containing both downloads and run:

python ./check-register.py ./lens-74-seven-day-check.csv

This invokes the same checker with the same result. Python is required; the PowerShell launcher itself has not been executed in this preparation environment.

Sources and evidence boundary

Source checked 3 November 2026: Ahrefs: How Stable Are AI Citations? Its observations are attributed to its own sample. The 3-question, 7-day register and decision procedure here are editorial controls, not a validated measurement standard.

No live buyer-question experiment, audience reach, measured conversion benefit or causal effect is claimed. Website readiness, mention, recommendation and citation are different observations.

About the publisher

VerisAI, founded in 2025, helps B2B services, SaaS, consulting and technology firms in Europe improve how their websites can be read, understood and cited by AI systems. Its AI Readiness Scoring & Monitoring service supports practical website checks. Contact support@verisai.eu.

Frequently asked questions

Does one missing AI citation mean our website is broken?

No. One missing citation is an observation. Check whether the response cited anyone, whether the tool completed the answer, and whether a repeat pattern exists before attributing the result to a website defect.

What should stay the same during the seven-day check?

Keep the exact question, AI product and available model, search setting, language, intended buyer market and available personalisation settings consistent. Use a fresh conversation for each run and record any condition you cannot control.

How do we calculate citation presence?

For each question and platform separately, divide captured answers citing the target website by all successfully captured answers under comparable settings. Include completed answers with no citations in this denominator. Report missed or failed checks separately.

Is a mention the same as a recommendation or a citation?

No. A mention names the brand. A recommendation suggests it as a suitable choice. A target-site citation links to its website as a source. Record all three separately; also note whether any source was cited.

What should we fix after a repeated gap?

Inspect the intended public page and the actual cited sources. Repair a confirmed access restriction, outdated or conflicting fact, or missing answer to the buyer question. Assign an owner and verify the specific repair. If no defect is established, continue observation.

Does seven days prove a trend or the effect of a page change?

No. Seven days is a manageable starting window. It does not establish statistical significance, represent all buyers or prove causation. A later citation can return without a website change; longer repeated observation may be needed.