Open-source AEO for your website: a practical 7-step guide

Arber Xhindoli · November 13, 2026 · 10 min read

Your website describes what you sell, but an AI answer recommends a competitor. Installing an AEO tool can help you investigate that gap. The useful result is a specific page change you can explain and measure.

You can implement open-source answer engine optimization on an existing website. Start with a small set of buyer questions, save the answers, inspect the pages behind them, and improve the missing information. Your CMS can stay in place. The measurement application can run separately.

I build Canonry, so this guide uses it for the worked setup. The process also gives you a way to evaluate other open-source AEO platforms: can the tool carry evidence from the first observation through to the next check?

01Measure

Save answers to real buyer questions.

02Investigate

Find the page or business fact that needs work.

03Improve

Review and publish a specific change.

04Check again

Verify the page and repeat the measurement.

A useful AEO workflow connects each website change to the question and evidence that prompted it.

1. Choose one buyer decision to improve

Start with a decision your website should help someone make. A hotel might focus on whether it suits families arriving by car. A software company might focus on whether its product supports a particular integration. A local contractor might focus on a service in one city.

Write down the questions a buyer asks before taking action. Include questions about fit, limits, alternatives, and cost. Keep the first group small enough that someone on your team can read every answer.

For an illustrative hotel project, that could include:

  • Which hotels near the station offer family rooms?
  • Where can I stay with an electric vehicle and overnight parking?
  • Which hotels allow late check-in after midnight?
  • What should I check before booking a room for two adults and two children?

These questions imply different facts. A generic paragraph about a convenient location cannot answer all of them. Map each question to the page that should provide the answer, such as rooms, parking, arrival information, or a useful local guide.

Keep branded questions separate. “Does Harbor Hotel have parking?” tests the description of a named business. “Which hotels near the station have parking?” tests whether the business enters a buyer's consideration set. Combining them makes your result harder to interpret.

Record the market and language too. A business with one domain may serve several distinct markets. If that describes your site, use the one-site, many-markets model before expanding the question list. Otherwise, a strong result in one city can hide a weak result in another.

The output of this step is a measurement brief: the buyer, decision, questions, relevant pages, and competitors. It gives every later change a purpose.

2. Install the platform separately from your website

Canonry is a self-hosted application. You can use it to inspect a public website without replacing that site's CMS or adding a new frontend. The software is MIT-licensed from version 5.0; hosting, provider usage, and your team's work still need a budget.

For a first local setup, the official quickstart documents these commands:

npm install -g @canonry/canonry
cnry bootstrap
cnry serve

Run them on your local computer with Node.js and npm installed. cnry serve runs in the foreground, so leave that terminal open while using it. Then visit http://127.0.0.1:4100/setup and follow the setup for your public site. The bootstrap output includes a credential; keep it private.

Start with the site audit. AI Visibility has a separate provider setup, so you can establish a technical baseline before configuring paid answer collection. Once that works, configure the provider you want to measure through the application settings. Review the provider's usage charges before collecting answers.

Before running a large site scan, set a page limit. Read the completion status and coverage when it finishes. If the scan reached 100 pages of a 2,000-page site, its findings describe those inspected pages. Expand the scan when you know which sections require coverage.

Use the deployment guide when moving beyond a local trial. Decide who owns updates, credentials, backups, and service availability. Canonry's standard self-hosted instance serves one operator or trusted team. Unrelated teams need separate instances; creating another project is not a substitute for that separation.

A successful installation gives you a working evidence store. It does not yet mean your website has become more visible.

3. Save a baseline you can compare later

An answer saying your brand's name and an answer linking to your website are different observations. Save both. Also preserve the exact question, answer text, source URLs, collection date, engine, and method.

Use a simple evidence record:

FieldWhat to recordWhy it matters
QuestionExact wording, market, and languageChanging the question changes the comparison
CollectionEngine or model, API or browser, dateDifferent collection methods can return different evidence
MentionWhether and how the brand appearsA name alone may not be a recommendation
CitationLinked source and destination pageShows which source supported the answer
CompetitorsBusinesses named in the same responseGives you a relevant comparison set
StatusCompleted, failed, or missing evidenceAn unsuccessful request is not a confirmed absence

Canonry keeps mentions and citations separate and preserves engine-specific results. Its provider documentation explains the available collection paths. When a report concerns the consumer ChatGPT experience, identify whether the observation actually came from that experience or from an API request. Label the result accordingly.

After provider setup, add your first questions and collect a run. In a second terminal, use the following pattern from the CLI reference, replacing my-site with the project name created during setup and the example questions with your own:

cnry query add my-site "Which hotels near the station offer family rooms?" "Which hotels allow check-in after midnight?"
cnry run my-site --wait
cnry visibility-stats my-site --by-provider

The run uses your configured providers and can incur usage charges. Check that it completed before interpreting the statistics, then open the saved answers and citations in the project. If a run is still active or failed, inspect its status before starting another one.

Choose a repeatable schedule that fits your budget and review capacity. Several comparable observations are more informative than a single screenshot. Keep exploratory questions outside the core group until you decide to track them regularly.

Also record what the website looks like before the change. Save the relevant page text or revision, its URL, and any known indexing or access issue. Without that record, a later improvement can be difficult to connect to the work you did.

This baseline is a sample of answers under stated conditions. It does not represent every answer every customer receives.

4. Fix access and page structure before adding more content

A useful answer on an inaccessible page cannot do its job. Inspect the pages mapped to your buyer questions and work through the concrete failures first.

Check that each public page loads, uses the intended URL, and exposes its important facts as readable text. Review robots rules, indexing directives, redirects, canonical URLs, and internal links. A page linked only from an old campaign email may be much harder to discover than one linked from the relevant service page.

Check the hosting layer as well. A public browser visit can work while a firewall or bot rule blocks another client. An audit finding should lead you to the actual response and rule involved, so you can make an intentional access decision.

Google's guidance for AI features says existing SEO fundamentals apply to AI Overviews and AI Mode. It does not require a special AI text file or special schema. That is guidance about Google's features; it should not be stretched into a claim about every answer engine.

Use structured data to describe information that is also visible and accurate. For the hotel example, an address in JSON-LD should agree with the address guests see. Fixing that inconsistency is valuable regardless of whether the next AI answer changes. See our structured-data guide for the role of page markup.

Connect Search Console if you need search performance or URL inspection evidence in the investigation. Keep Google indexing status separate from the question of whether a particular AI service cited the page.

Finish this step with a small repair list: the affected URL, the observed failure, the proposed change, and the person who can implement it.

5. Write the answer your buyer still needs

Now compare the buyer question with the page. What would a reader still need to know before making the decision?

Suppose the parking page says only, “Parking available.” The missing information might be the overnight charge, vehicle height limit, charging availability, booking requirements, and what happens when the garage is full. Those are useful facts to obtain from the business before writing more copy.

Put the direct answer near the relevant heading. Follow it with the details and exceptions that affect a customer's choice. A short explanation of where to park an oversized vehicle can be more useful than another page repeating that the hotel is centrally located.

Choose the format from the question:

  • Use a listicle when readers need several distinct options or checks.
  • Use a comparison table when the same attributes apply to each option.
  • Use a step-by-step guide when order matters.
  • Use a focused service or location page when the reader needs to decide whether to contact that business.

Use the sources you control to make the answer accurate: service specifications, staff input, current policies, product documentation, or a clearly described customer example. Link to supporting material where readers can verify an important claim.

Review external business information when it conflicts with the site. A current page and an outdated public profile can tell different stories. Correct the sources you manage and request corrections from relevant publishers where appropriate.

Keep each page's purpose clear. If three articles answer the same question in almost the same way, improving one strong page may be more useful than maintaining all three. The goal is a useful answer at a URL your team can keep current.

6. Publish a bounded change and verify the live page

Choose one page or a small related group for the first implementation. Save the old version and document what you expect the change to resolve. That makes review easier and gives you something concrete to inspect after release.

The publishing path depends on your website. A developer can prepare a pull request. A content team can use its CMS review process. Canonry's WordPress integration supports page reads and writes through the REST API. Its documentation distinguishes those operations from manual tasks such as schema injection and some staging work.

Review the actual page before approving publication. Confirm the business facts, links, headings, and any structured data. If an agent helped draft the change, give it a specific task and an explicit publishing boundary.

For example, an editorial task might be:

Draft an update to the parking page using the approved parking policy. Answer the height-limit and late-arrival questions. Preserve the booking link and send the draft for review.

After publication, open the public URL and verify the change arrived. Check that the page still loads, that the booking or contact action works, and that the original audit issue is resolved. Record the release date and revision.

These are the first success criteria you can verify directly. An AI citation is a separate outcome to measure afterward.

7. Repeat the checks and decide what to do next

Rerun the technical check on the changed page. Then repeat the tracked buyer questions with comparable settings and review the saved answers. Keep new experiments separate from the baseline until you decide how to include them.

Compare three kinds of evidence. The page check shows whether you shipped the intended improvement. Answer records show whether sampled mentions or citations changed. Traffic and business data show whether measurable visits and useful actions followed.

A crawler request alone does not establish a citation, and a citation does not establish a booking. Use server traffic evidence, referrals, and your own conversion records for the questions each can answer. If a measurement failed or a source is unavailable, leave the gap visible.

For an illustrative first month, use the first week to choose questions and establish a baseline. Use the next two weeks to investigate and publish one manageable change. Use the remaining time to verify implementation and review repeated measurements. This is a work schedule, not a promise that AI results will move within a month.

At the review, decide whether to keep observing, fix another problem, or expand the question group. Record the reason. A workflow becomes useful when evidence changes the team's next action.

Start with the open-source platform if your team can own setup and implementation. Use the resource guide for the documentation behind each task. If you want an engineering team to handle the website and ongoing AEO work, Canonry Managed provides a scoped delivery path.

Can I use open-source AEO on an existing website?

Yes. You can run measurement and audits outside your website, then make changes through your existing CMS or codebase. Canonry runs as a separate self-hosted application. Publishing connections, such as its WordPress integration, are optional and have their own setup.

Is open-source AEO free?

Canonry 5.0 and later have no software license fee. You still need to account for hosting, maintenance, implementation time, and any paid providers you connect. A technical site audit and an AI-answer measurement run are different jobs; provider-backed answer collection has its own usage costs.

How do I know whether my AEO changes worked?

First verify that the intended page change is live and that the original technical or content issue is resolved. Then repeat a stable set of buyer questions using comparable collection settings. Review saved mentions and citations alongside referrals and business outcomes. A higher audit score or one favorable answer alone does not establish a lasting visibility improvement.

Next

Continue with the platform.

Inspect the technical workflow, run it on your own site, or add live visibility reporting to an agency portal.

Open Canonry platform