How to check whether AI answers cite your website

To check whether AI answers cite your website, run a fixed set of customer questions and save the answers with their source URLs. Count a citation only when the returned source points to your site. A mention of your company and a link to your domain are separate observations.

Start with a small question set you can inspect yourself. The result should explain where your site appeared in those answers, with enough detail to repeat the check. It cannot establish everything an engine says about your business.

Write the question set before looking at results

Choose questions that reflect decisions your customers face. Include an unbranded category question, a specific use case and a comparison question. Keep explicitly branded questions in their own group; asking about your company makes a mention more likely by construction.

After connecting Vaaya, use this prompt:

Check these exact questions: [questions].
Target website: [domain]. Country: [country]. Engines: [chosen engines].
Keep each answer and every returned citation with the observation time.
Distinguish a brand mention, a source on our domain and a recommendation.
Show failed or missing answers separately. Do not rewrite questions mid-run.
Stay within [total budget] and set a maximum on each paid call.
Return the evidence table before suggesting content changes.

Avoid feeding the agent a list of pages you hope it finds unless that is the experiment you intend to run. Save the original question list as a versioned input.

Run one bounded batch per engine

1. Discover the current route. Vaaya exposes olostep/ai-visibility. The current schema accepts engine, an array of queries, and optional country. Accepted engine values are chatgpt, perplexity, gemini, copilot, google-ai-mode and google-ai-overview. Confirm availability and the quote before submitting.

2. Submit the fixed questions. One request selects one engine. Reuse the same question array for the next engine so the comparison has a common input. Ask for current pricing for the complete batch, including any separate retrieval steps.

3. Keep the returned batch ID. This action uses Olostep's batch flow. Poll olostep/batch-status with the returned id; after completion retrieve olostep/batch-items with the same id and formats: ["json"]. Parse each item's json_content string to read the structured answer and sources. Do not submit the questions again to check progress. Follow the returned workflow if the current contract changes.

4. Inspect answers and source links. Preserve answer text, source URLs and inline references where supplied. Olostep's AI visibility product describes this kind of answer and citation capture. Read a few source pages to check that the cited passage supports the associated claim.

5. Classify the observation. Record brand mentioned, target domain cited, recommendation present and answer missing as separate fields. Let a reviewer resolve ambiguous cases before counting them. A list of tools is not necessarily an endorsement of every item.

Make the denominator visible

Use a table with this shape:

run_id | prompt_version | prompt | engine | country | observed_at
answer_status | answer_text | brand_mentioned | target_domain_cited
cited_urls | recommendation_excerpt | reviewer_note

Normalize hostnames carefully. Decide before scoring whether a help subdomain counts as your website and how redirected URLs should be handled. Preserve the original URL alongside the normalized value so the choice can be audited.

Report successful answers and missing answers separately. If two of ten responses failed, a citation share among eight completed answers must say so. Counting failures as uncited answers confuses availability with content; silently removing them hides part of the run.

Turn a source gap into a specific edit

Open the pages the engine cited. Check whether your own corresponding page states the relevant fact plainly, supports it with evidence and is accessible. That may suggest a documentation edit or a new example. It does not establish why the engine selected another page.

Repeat the same question set after an agreed interval. Your application owns the schedule, run storage and cumulative budget. Keep extra exploratory questions in a separate run so they do not change the original comparison.

End the review with a small set of proposed edits, each tied to an observed answer and a missing source fact. Recheck after those edits using the saved prompts and preserve both runs, including results that moved in the opposite direction.

Questions

Does a mention count as a citation?

No. Record a brand mention and a linked source separately. An answer can mention a product while citing another website.

Will the same prompt always produce the same sources?

No. Preserve the prompt, engine, country, time and actual answer for each observation. Repeated runs may differ.

Does Vaaya schedule the citation check automatically?

The workflow described here runs a bounded check. Your application or scheduler owns recurring runs, storage and notifications.