Portfolio

How do you build a GEO portfolio project that proves your skills?

Written by Albrecht Lippky

Last updated: • 12 min read · Markdown

A credible GEO portfolio project uses ten real customer questions, three assistants, repeated runs and one documented content or technical change.

An open graph-paper notebook with a hand-drawn tally grid and rising line, a highlighter and printed pages
Jump to section
  1. What is a GEO portfolio project?
  2. What should the project prove?
  3. What is the recommended GEO portfolio brief?
  4. How do you choose a brand and get permission?
  5. How do you build a useful prompt set?
  6. How do you record the baseline?
  7. How do you define the metrics?
  8. What single change should you make?
  9. How should you measure over the following weeks?
  10. How can you measure without overstating causality?
  11. Which tools should you use?
  12. How should you write the GEO case study?
  13. Which GEO portfolio mistakes should you avoid?
  14. FAQ
  15. Sources

Summarize with ChatGPT Claude Perplexity

TL;DR

  • A credible GEO portfolio project uses ten real customer questions, three assistants, repeated runs and one documented content or technical change.[1]
  • Keep brand mentions, domain citations, Google generative-AI impressions, AI referrals and conversions as separate metrics.[3][5]
  • Use a spreadsheet first, then add automation only when its prompt allowance and collection cadence match the project protocol.[1]
  • Hiring evidence is stronger when employers can inspect exact prompts, dated before-and-after work, measurement choices and lessons learned.[5]
  • Describe observed changes honestly because variable assistant answers and measurement noise can make a before-and-after result non-causal.[1]

A strong GEO portfolio project does not promise that one page “ranked number one in AI.” It shows that you selected meaningful customer questions, recorded repeatable observations, made one bounded change, measured over time and explained what the evidence can—and cannot—prove.[1]

For learners, the most credible format is a small, inspectable case study:

  • One owned site, employer-approved brand or consenting non-profit.
  • Ten real customer questions.
  • Three AI assistants.
  • Three repeated runs at each measurement point.
  • One documented content or technical change.
  • Four weeks of follow-up measurement.
  • Raw answers, cited URLs, timestamps and limitations.

That package can demonstrate practical GEO, AEO and AI-search skills beyond a course certificate without overstating causality.

What is a GEO portfolio project?

A GEO portfolio project is a documented experiment showing how you investigate and improve a brand’s visibility in generative search. It combines prompt research, AI-answer analysis, content or technical implementation, repeated measurement and written interpretation.

The project should distinguish between several outcomes:

  • Brand mention: whether the brand appears in an assistant’s answer.
  • Domain citation: whether the assistant links to the brand’s website.
  • Page citation: which specific page is linked.
  • Position: where the brand appears among mentioned entities.
  • Search impression: whether Google shows a link to the site in AI Overviews or AI Mode.
  • Referral visit: whether a user arrives from an AI assistant.
  • Conversion: whether the visit produces a measurable business action.

Google’s generative-AI performance report measures link impressions in AI Overviews and AI Mode; that is not the same as counting textual brand mentions.[2][3] ChatGPT referrals can be identified through inbound visits using utm_source=chatgpt.com, while assistant citations and visits remain separate observations.[4][1]

Do not combine these signals into one unexplained “GEO improvement” score. A portfolio becomes more credible when it defines each metric and preserves the denominator behind it.[1]

What should the project prove?

The project should prove that you can perform a controlled, honest workflow—not that every AI answer changed because of your work.

A hiring manager should be able to verify:

  • Why the brand and audience were selected.
  • That you had permission to work on the site.
  • Where the questions came from.
  • Which assistants and search settings were used.
  • What the baseline answers contained.
  • What exactly changed on the site.
  • When the change was published.
  • How the same questions were re-tested.
  • Which results were positive, unchanged or negative.
  • What alternative explanations remain.

This matches the evidence requested in several 2026 hiring examples. Career Certified asks candidates to explain what they tested, measured and learned, while Intrepid Digital emphasizes visibility, citations, sentiment, accuracy and business outcomes.[3] One GEO/AEO agency listing asks for exact prompts, platforms and dated before-and-after evidence.[5]

These postings do not establish a universal portfolio requirement, but they indicate the type of work that is easier for employers to inspect.

Use the following assignment as a practical five-week project. Its sample size is a suggested learning design, not a research-validated minimum.

Step Output Time Tool options
Scope the project Permission record, audience, market and target page 1–2 hours Brief document, email approval
Build the prompt panel Ten customer questions with sources and intent labels 2–4 hours Spreadsheet, support tickets, interviews, sales notes
Capture the baseline Ninety saved answers across three assistants 1–2 days ChatGPT Search, Perplexity, Gemini, spreadsheet
Define one change Intervention plan and success metrics 1–2 hours Content editor, CMS, technical audit
Implement the change Before/after copy or configuration diff 1–3 days CMS, Git, browser tools, Search Console
Measure weekly Four follow-up waves of ninety answers About 1 hour per week Same assistants, spreadsheet or optional tracker
Add supporting data Search impressions, referrals or conversions where available 1–3 hours Search Console, analytics platform, server records
Write the case study Report, evidence folder and limitations 1–2 days Markdown, document editor, slide deck

The baseline contains 10 × 3 × 3 = 90 answers: ten prompts, three assistants and three runs per prompt-assistant pair. Four subsequent weekly waves produce another 360 answers, for 450 saved answers across the complete assignment.

The numbers make the project concrete, but they do not make it representative of every customer question. Report both the total number of answers and the smaller prompt-level denominator.

How do you choose a brand and get permission?

Choose one of three safe options:

  1. Your own website: Use a site where you control the content and can show the implementation.
  2. An employer’s site: Obtain written permission and clarify whether the work may be published.
  3. A non-profit: Select an organization that agrees to the scope, access and publication terms.

Avoid using a recognizable company without authorization. If the work is confidential, anonymize the brand and clearly label the case study as anonymized. Do not present a self-directed project as client work, and do not invent a sponsor, testimonial or reference.[6][7]

Record:

  • Organization and website.
  • Your role.
  • Permission date.
  • Approved pages and markets.
  • Whether screenshots, URLs and raw assistant answers may be published.
  • Any confidential details removed from the prompt set.

How do you build a useful prompt set?

Start with questions real customers might ask before choosing a product, service or organization. The strongest prompts are not generic keyword variations; they represent decisions, concerns and comparisons.

Good sources include:

  • Customer-support questions.
  • Sales-call notes.
  • Product reviews.
  • Frequently asked questions.
  • Search Console queries.
  • Website search logs.
  • Public forum discussions.
  • Interviews conducted with permission.
  • People Also Ask questions.
  • Existing content gaps.

Remove names, email addresses, order numbers and other identifying details before adding questions to the project.

Create a spreadsheet with these columns:

Column What to record
Prompt ID A stable identifier such as P01
Verbatim question The exact wording tested
Source Support, sales, interview, search data or another source
Intent Informational, comparison, problem-solving or transactional
Audience The relevant customer segment
Funnel stage Discovery, consideration, decision or post-purchase
Expected evidence The page or information that should answer it
Selection rationale Why this question matters
Market and language Country, language and location settings

Keep the ten questions fixed throughout the project. Changing the prompts midway makes the comparison harder to interpret.

How do you record the baseline?

Use three assistants, such as ChatGPT Search, Perplexity and Gemini. If you also test Google AI Overviews, label it as a separate search surface rather than silently treating it as another Gemini result. Finseo’s own analysis found that Google AI Mode and AI Overviews share only 13.7% of cited URLs on the same query, so give each Google surface its own column as well.

For every run, save:

  • Prompt text.
  • Assistant and model name as displayed.
  • Search mode or web-search setting.
  • Date and time.
  • Country, language and device conditions where relevant.
  • Account or personalization conditions that may matter.
  • Full response.
  • Cited URLs.
  • Brand mention status.
  • Owned-domain citation status.
  • Specific page cited.
  • Approximate brand position.
  • Run ID.
  • Screenshot or exported capture.

OpenAI warns that search results and citations can be incomplete, outdated or incorrect.[8] Check whether cited pages actually support the assistant’s statement instead of assuming that a citation proves accuracy.

Treat failed runs as failures. Do not automatically code a technical error, unavailable search result or truncated answer as a non-mention.

How do you define the metrics?

Use simple, visible definitions.

Mention rate is the number of responses mentioning the brand divided by the number of valid responses.

Citation prevalence is the number of responses citing the brand’s domain at least once divided by the number of valid responses.[1]

Citation share compares the brand’s citations with all citations collected for the same question set. It answers a different question from citation prevalence.[1]

Position records where the brand first appears among named brands. This matters because being included and being recommended first are different outcomes. SparkToro’s City of Hope experiment illustrates that distinction: the organization appeared in 69 of 71 ChatGPT answers but appeared first in only 25.[6]

Search impression records whether Google showed a link to the site in a generative-AI feature. Google states that its dedicated report can break this data down by page, country, date and device, although small sites may not see the report if they have insufficient generative-AI impressions.[9]

Referral traffic records visits attributed to an AI assistant. Conversions record business outcomes only when the project has a suitable analytics or transaction system. Treat these counts cautiously: Finseo’s AI search attribution study reports that analytics records only about 1 in 10 AI-driven conversions.

Keep separate columns for each metric. A rise in citations is not automatically a rise in visits, and a rise in visits is not automatically a rise in revenue.

What single change should you make?

Choose one bounded intervention after reviewing the baseline. Two suitable routes are:

  • Rewrite one service or product page to answer a documented customer question with original, attributable information.
  • Correct one verified crawlability, indexing or technical obstacle.

Google recommends useful, original content and first-hand knowledge for generative-AI search features. Pages must also be indexed and eligible for a Search snippet.[10]

Record:

  • The baseline page or configuration.
  • The problem identified.
  • The hypothesis.
  • The exact change.
  • Publication or deployment timestamp.
  • Screenshots or version history.
  • Any related changes made by other people.
  • The metric expected to move.
  • The metrics that are explicitly not being tested.

Do not make “add special AI schema” the default intervention. Google says special AI text files and special schema markup are not required for its generative-AI search features, and eligibility still does not guarantee crawling, indexing or serving.[10]

A single intervention keeps the case study understandable. If you rewrite a page, change internal links, acquire backlinks and alter technical settings at the same time, you cannot clearly explain which action relates to the observed result.

How should you measure over the following weeks?

Repeat the exact prompt panel once per week for four weeks. Use the same assistants, settings and capture procedure whenever possible.

Maintain a measurement log containing:

  • Collection date.
  • Completed and failed runs.
  • Prompt-level results.
  • Assistant-level results.
  • New or removed citations.
  • Position changes.
  • Page changes.
  • Search Console observations.
  • AI referral visits.
  • Unrelated marketing or technical changes.
  • Notes about unusual answers.

Repeated answers can vary even when the prompt does not change. One academic study found that 9%–27% of questions changed their overall answer decision within five minutes under its tested conditions.[11] Another study reported retrieved-link overlap of 18% for Google AI Overviews across measurements roughly two months apart, compared with a highest overlap of 45% for organic search.[11]

These findings do not provide a universal volatility rate for your project. They show why preserving repeated observations is more credible than comparing two screenshots.

How can you measure without overstating causality?

How should a learner explain a before-and-after change?

Use language that describes the observation rather than claiming an untested cause. For example: “After the page revision, the brand was cited in 8 of 30 valid assistant responses, compared with 4 of 30 at baseline.” Then state that the project cannot isolate the page revision from answer variability, index changes, competitor changes or other site activity.

A defensible conclusion might say:

“During the five-week observation period, citation prevalence increased from 4/30 to 8/30 for the selected assistant panel after the page revision. The result is directional rather than causal because the project had no control group, the prompt set was small and assistant answers varied between runs.”

Avoid:

  • “The change caused a 100% increase in GEO.”
  • “The brand now ranks number one in AI.”
  • “This proves the page is preferred by ChatGPT.”
  • “The certificate demonstrates job readiness.”
  • “The site became visible everywhere.”

Academic work shows that answer decisions and cited sources can change independently.[11] A 2026 statistical framework also found that many apparent domain differences may fall within measurement noise and did not establish a universal sample size for this assignment design.[1]

If the result does not improve, report that clearly. A well-documented null or negative result can demonstrate experimental discipline, especially when it includes what you learned and what you would test next.[3]

Which tools should you use?

Start with a spreadsheet. Manual capture gives you control over the questions, raw-answer archive and scoring definitions. It also prevents a dashboard from hiding the underlying evidence.

Optional tools can reduce collection work:

  • Finseo: Its free AI visibility check is available without login, while its Starter plan offers a seven-day trial with 50 tracked prompts, three models and daily or weekly tracking.[12][13] For employer-approved projects, its enterprise page lists SSO via SAML, roles per project, an audit trail, unlimited seats, hosting in Frankfurt, Germany with EU data residency, GDPR compliance with a data processing agreement (DPA) and ISO 27001 certification.
  • Ahrefs: Its free AI Overviews Tracker provides a no-signup visibility preview; custom prompts are available through qualifying paid plans.[2]
  • Rank Prompt: Its free plan provides ten credits for one report using the user’s own prompts; recurring free capacity is not published.[14][1]
  • Beaconn: Its free plan provides one scan covering six buyer questions across ChatGPT, Perplexity, Gemini and Claude; recurring free refresh capacity is not published.[15]
  • Otterly: Its November 2026 announcement lists Lite at $29 per month with 20 daily-tracked prompts across three engines, or 16 prompts across four engines, plus 2,000 monthly credits.[12]

A free scan, a free trial and recurring free monitoring are different offers. Check the collection cadence, engine coverage, prompt allowance and export options before selecting automation.

Keep the original assistant answers even when a tracker is used. The portfolio should be inspectable without requiring a hiring manager to trust a proprietary score.

How should you write the GEO case study?

Use this short outline:

  1. Project summary

    • Brand, audience, market and project dates.
    • Your role and permission status.
    • One-sentence objective.
  2. Question and hypothesis

    • Customer problem investigated.
    • Ten-question panel.
    • Expected effect of the intervention.
  3. Baseline method

    • Assistants tested.
    • Number of prompts and runs.
    • Search settings and collection dates.
    • Metric definitions.
  4. Baseline findings

    • Mention counts.
    • Citation counts.
    • Position observations.
    • Important answer excerpts or linked evidence.
  5. Intervention

    • Page or technical element changed.
    • Before/after evidence.
    • Publication date.
    • Reason for selecting the change.
  6. Follow-up results

    • Weekly table.
    • Valid and failed runs.
    • Assistant-level and prompt-level changes.
    • Search impressions, referrals or conversions where measured.
  7. Interpretation

    • What changed.
    • What did not change.
    • What cannot be attributed to the intervention.
    • Alternative explanations.
  8. Next test

    • One follow-up experiment.
    • New evidence required.
    • Decision rule for continuing or stopping.
  9. Evidence appendix

    • Raw-answer log.
    • Prompt-source register.
    • Screenshots.
    • Cited URLs.
    • Copy or configuration diff.
    • Permission statement.

Use a dated evidence folder and stable filenames such as P01_chatgpt_week1_run2. Link to the live page when permitted, but do not publish confidential prompts or personal information.

Which GEO portfolio mistakes should you avoid?

  • Testing only one prompt once.
  • Changing the prompt set after seeing the results.
  • Treating a citation as proof that the cited page supports the claim.
  • Calling a crawler visit a customer visit.
  • Combining mentions, citations, impressions and conversions into one score.
  • Claiming causality from two screenshots.
  • Making several major changes at the same time.
  • Using an employer’s data without written permission.
  • Describing a free check as permanent free monitoring.
  • Reporting only positive results.
  • Hiding failed runs or unavailable data.
  • Claiming that a certificate alone proves practical ability.

OpenAI distinguishes OAI-SearchBot, used for search discovery, from GPTBot, associated with model training.[16] Perplexity similarly describes its crawler as helping surface and link websites rather than collecting foundation-model training content.[17] Crawler access belongs in the implementation record, not in the results column.

Frequently asked questions

Is a GEO portfolio project better than a GEO certificate?

A certificate can show that you completed structured learning, while a portfolio project can show how you applied that learning. A credible project adds inspectable evidence: prompts, raw answers, implementation records, repeated measurements and limitations. Career Certified’s hiring language specifically emphasizes what candidates tested, measured and learned.[3]

How many prompts should an AEO portfolio project include?

Ten fixed customer questions are a practical scope for a learner project. Test each question three times across three assistants at baseline and repeat the panel weekly for four additional waves. This creates a clear denominator, but it is not a research-validated minimum sample size.[1]

Can I build a GEO case study for my own website?

Yes, an owned website is one of the safest project choices because you can document the change and link to the result. State that it is self-directed work, define your role and avoid presenting it as client work. If the site belongs to an employer or non-profit, obtain permission before publishing details.

Which AI assistants should I test?

ChatGPT Search, Perplexity and Gemini provide a manageable three-assistant panel. Google AI Overviews can be included, but label it as a separate search surface rather than silently treating it as Gemini. Keep the assistants and settings consistent throughout the project.

Can I claim that my content change caused better AI visibility?

Usually not from a small before-and-after project. Assistant answers and cited sources can vary over short periods, and apparent differences may fall within measurement noise.[1] Report the observed change, denominator, timing and limitations instead of presenting correlation as proof of causation.

Should I use a paid GEO tracking tool?

Not necessarily. A spreadsheet and manual captures are enough to complete the recommended assignment and preserve raw evidence. Paid tools can help with recurring collection, but compare their engine coverage, prompt limits, cadence and exports; Finseo’s Starter access, for example, is presented as a seven-day trial rather than a permanent free monitoring tier.[1]

Sources

  1. Quantifying Uncertainty in AI Visibility A Statistical Framework for Generative Search Measurement · arxiv.org
  2. Free AI Overviews Tracker by Ahrefs: Track Your Brand in AI Overviews · ahrefs.com
  3. SEO Manager, AI Search & GEO at Career Certified (Remote US (residency required)) · seojobs.com
  4. Publishers and Developers - FAQ | OpenAI Help Center · help.openai.com
  5. GEO / AEO Specialist (Full Time, Freelance Contract) · upwork.com
  6. NEW Research: AIs are highly inconsistent when recommending brands or products; marketers should take care when tracking AI visibility - SparkToro · sparktoro.com
  7. GEO Manager (Generative Engine Optimization) | Intrepid Digital · gointrepid.com
  8. Searching the web with ChatGPT | OpenAI Help Center · help.openai.com
  9. Generative AI performance report (Search) - Search Console Help · support.google.com
  10. Google's Guide to Optimizing for Generative AI Features on Google Search | Google Search Central  |  Documentation  |  Google for Developers · developers.google.com
  11. Characterizing Web Search in the Age of Generative AI · aclanthology.org
  12. OtterlyAI: AI Search Analytics From $29 a Month, Now With · globenewswire.com
  13. Pricing, Starter, Business, Growth & Enterprise | Finseo · finseo.ai
  14. Rank Prompt Pricing: AI Visibility Plans, Free to Start · rankprompt.com
  15. Pricing · AI Visibility Tracking from $0 | Beaconn · beaconn.ai
  16. Overview of OpenAI Crawlers · developers.openai.com
  17. Perplexity Crawlers · docs.perplexity.ai

More from GEO Courses

All guides