GEO on a Budget: A Practical AI Search Pilot for Small Teams
Start GEO with a focused website pilot. Improve five pages, measure AI citations and qualified leads, and assess ROI before increasing your budget.
Ryan Wu
11/18/202610 min read
How to improve five website pages, test AI visibility, and measure the business case before increasing your spend.
You search for your industry in an AI assistant. Competitors appear in the answer. Your business does not. Then someone asks the marketing team: “Should we be doing GEO?”
The concern is understandable. The next step is less obvious. Who owns the work? Does the website need rebuilding? Should you buy a tool or hire an agency? And what would count as a worthwhile return?
My recommendation is to start with a tightly scoped website pilot. Improve three to five commercially important pages, resolve access problems, publish the information buyers need, and monitor a consistent set of customer questions.
Even if AI citations do not follow immediately, that work can leave you with better product information and more useful sales resources. The first investment should produce something you can inspect and use.
This approach assumes you already have a website, permission to edit it, and access to colleagues in IT, marketing, and sales or product management. “Low budget” means limiting additional purchases and outsourcing. Internal time still needs a budget.
Before you start: three decisions
Is this the right priority? A pilot makes sense when customers need to compare specifications, evaluate alternatives, or verify credentials before buying—and your website does not explain those things clearly. That is particularly relevant for complex products and long sales cycles. If your positioning is changing every week, you cannot edit the site, or you have no approved product information to publish, resolve those constraints first.
Do you need a new website? Inspect the current one before deciding. If it serves readable content reliably and allows page edits, targeted improvements may be enough. A five-page pilot should have a clear boundary.
When should you expect results? Separate delivery dates from visibility outcomes. Your team can commit to completing page improvements. It cannot set a date when an external platform will cite them.
Generative engine optimization, or GEO, aims to improve how accurately and visibly a business appears in AI-generated answers. There is no universal AI recommendation directory that guarantees inclusion after you submit a URL. Accessibility, indexing, citation, and recommendation are different outcomes. Begin with what your team can control.
1. Give IT a specific access checklist
A page that opens in your browser may still be difficult for automated systems to retrieve. Login requirements, verification challenges, firewall rules, redirect problems, and content that depends on complex interactions can all deserve investigation.
Ask IT to check four things:
Page availability: Do the selected URLs return the expected pages without access failures, login barriers, or unexpected redirects?
Crawling and indexing controls: Do robots.txt rules or page-level directives unintentionally restrict public content?
Traffic filtering: Do CDN, firewall, or server logs show legitimate search traffic being blocked?
Readable product information: Can the main description and specifications be retrieved reliably, or are they available only inside images, attachments, or interactive components?
Adjust access rules using the target platform’s official documentation. Keep security protections in place and distinguish search crawling, model training, and user-initiated retrieval.
For example, OpenAI describes OAI-SearchBot as its search crawler, GPTBot as a crawler for content that may be used in model training, and ChatGPT-User as a user-initiated agent. Allowing training access is not a guarantee of search visibility. OpenAI crawler documentation
Acceptance criteria: A documented issue list, a record of fixes, and repeatable checks showing that the selected pages can be accessed and read. Pages that already work do not need an architectural overhaul to satisfy the project.
2. Create one verified source of company and product facts
The homepage uses a short brand name. The brochure uses the legal company name. An old announcement uses a previous name. Two pages list different specifications for the same model.
Publishing more content in that situation multiplies the maintenance problem. First, build a shared reference containing:
The relationship between the company, brand, and product names.
Official contact details and product model identifiers.
Verified specifications, including units and operating conditions.
Certifications, supporting reports, and their scope.
Document versions, review dates, and the person responsible for each fact.
Marketing can organize the information; product or sales colleagues should verify it. The website team can then update pages and add appropriate structured data, such as Organization, Product, or BreadcrumbList, where applicable.
Structured data expresses existing facts in a consistent format. Google says its AI search features do not require special schema markup. Treat implementation as part of good website maintenance, with no promise of inclusion. Google Search Central guidance
Acceptance criteria: The reference document, web pages, attachments, and structured data agree. Claims about certifications, customer results, and reviews have evidence behind them.
3. Improve the pages closest to a buying decision
Choose three to five pages based on commercial usefulness: product pages salespeople already share, priority solution pages, and pages that should answer questions customers repeatedly ask.
Each page should answer five questions:
What is this product or solution?
Who is it suitable for?
What are its important specifications?
What conditions or limitations apply?
What evidence supports the claims?
“Energy-efficient and suitable for a wide range of applications” gives a buyer little to evaluate. A more useful structure is:
[Product model] is designed for [specific application] and addresses [specific problem]. Its key specifications are [verified figures and units]. It requires [operating or system conditions]. Suitability under [exceptional condition] needs further assessment. Supporting evidence: [report title, date, and link].
This is a writing template. Replace every placeholder with verified information before publishing.
Put essential specifications in the page body and readable HTML tables. A visitor should be able to find a basic operating range without downloading an entire manual. For scanned PDFs, an OCR text layer can help, but someone must check model names, units, and numbers. Keep the core facts on the web page and use attachments for detail.
Claims about lower costs or higher efficiency need a comparison, test conditions, and a source. When evidence is limited, narrow the claim. Explain unsuitable applications too: boundaries help serious buyers assess fit.
Acceptance criteria: A prospective customer can understand the main application, specifications, requirements, and evidence without contacting sales. Sales colleagues can use the page for initial qualification and routine questions.
Once these pages are useful, expand into selection guides, comparisons, and implementation-cost explainers. Draw topics from sales notes and customer conversations: Which specifications matter first? When does option A make more sense than option B? What installation and maintenance costs sit beyond the purchase price?
Comparisons should disclose their conditions and sources, including the limits of your own product. A self-published ranking should be clearly identified as such.
4. Establish a baseline you can repeat
A screenshot showing an AI assistant mentioning your company is a useful observation. It is not a complete performance report.
Choose the platforms your customers actually use and create a small set of 10–20 realistic questions. Separate branded questions from unbranded category and selection questions. Asking about your company by name and counting the resulting mention as an unsolicited recommendation gives a misleading picture.
For each observation, record the date, exact prompt, platform and mode, answer, cited sources, and factual errors. Use fresh conversations and comparable web-search settings where possible. Repeat observations over time.
Track brand mentions and citations to your own website separately. Report the denominator: “Our website was cited in 3 of 15 tested answers” is clearer than an unexplained percentage. A small, fixed question set is a diagnostic sample, not a measure of platform-wide market share.
Then connect those observations to identifiable referral visits and qualified inquiries. Ask prospects how they first heard about you. Expect gaps in source data, and keep unrelated organic growth out of the GEO total.
A two-week pilot brief
The two weeks below are an internal delivery schedule. Start the clock once access, information, and people are available; it is not an indexing or citation promise.
Scope: Three to five priority pages and 10–20 customer questions, focused on one product area.
Owner: One marketing or business lead coordinates the project. IT checks access, a website developer handles template changes, marketing prepares content, and product or sales verifies claims. People can hold multiple roles, but someone must own acceptance.
Week one: Capture the baseline, inspect access, assemble the verified information, and agree on page changes.
Week two: Complete edits, conduct the business review, check the published pages, and establish a weekly observation log.
Budget record: Log hours by role and any development, software, outsourcing, or content-production expenses.
Deliverables: A page list, change log, supporting evidence, acceptance results, and baseline observations.
If a technical issue exceeds internal capacity, purchase help for that defined problem—for example, making product specifications readable or updating a product template. A precise scope and acceptance criteria make external work easier to manage.
What if citations still do not appear?
Investigate in order. Are the pages accessible? Are crawl or indexing records available? Does the content actually answer the tested questions? What useful evidence do the sources currently being cited provide?
Schedule a review four to six weeks after launch as a management checkpoint. If access problems remain, fix them. If the content misses the question, revise it. If the evidence is weak, collect stronger material before commissioning a large batch of new articles.
Further investment should follow a specific finding and a plausible next step. The absence of results alone is not a reason to publish more.
Measure visibility, opportunities, and profit separately
GEO reporting becomes confusing when marketing presents citations and management expects orders. Agree at the outset on costs, qualified leads, and attribution rules.
Visibility includes mentions, website citations, and information accuracy within your test sample. These help diagnose performance; they are not revenue.
Opportunities include traceable visits, qualified inquiries, sales-accepted opportunities, and movement into proposals or quotations. Define a qualified inquiry before counting it—for example, a reachable prospect in the target market with a relevant need. Exclude spam, job applications, and supplier pitches.
Profit means the contribution profit allocated to GEO under your agreed attribution method. Large contract values do not automatically imply attractive project economics.
Include the full cost of the work
Use a consistent calculation:
Total GEO cost = internal hours × agreed hourly cost + development and outsourcing + tools and materials + ongoing maintenance.
Separate initial improvements from recurring operations, and cash spending from the full cost including staff time. When a website improvement serves SEO, sales, and GEO, agree on a reasonable allocation in advance and retain the full project cost for comparison. Apply the same discipline to both costs and benefits.
Preserve source evidence and agree on attribution
Connect website source records, form submissions, and sales notes. Distinguish leads with recorded AI referrals, customer-reported AI discovery, evidence of multiple channel influences, and unknown sources. Customer recollection adds context but is not precise tracking.
Deduplicate customers across channels. The attribution shares for an order should add up to no more than 100%. Choose your reporting method before sales close, whether it uses first contact, last contact, or a documented allocation across touchpoints. Google’s attribution documentation explains the underlying principle: assigning conversion credit across interactions. Google Analytics attribution overview
Attribution does not establish incremental impact. An AI referral cannot prove the customer would otherwise never have purchased. Comparing similar pages before and after changes—or retaining comparable unchanged pages as a reference—can provide context. Record advertising, promotions, and seasonal changes alongside those comparisons. These observations still have limits as evidence of causation.
Calculate ROI using contribution profit
Attributed ROI = (attributed contribution profit − total GEO cost) ÷ total GEO cost × 100%.
For this calculation, contribution profit means revenue less product, delivery, and relevant variable costs, before subtracting the GEO spending shown separately in the formula. Align the definition with finance to avoid double-counting costs. If you use gross profit instead, label that basis explicitly.
Consider a hypothetical pilot costing CNY 12,000 over a quarter, including staff time, development, and maintenance. Leads acquired during that period eventually produce two orders, each contributing CNY 10,000 in profit before GEO costs.
At 100% attribution to GEO, attributed contribution profit is CNY 20,000 and ROI is approximately 66.7%.
At a documented 50% attribution share, attributed contribution profit is CNY 10,000 and ROI is approximately −16.7%.
The same orders lead to different conclusions under different attribution assumptions. Show the evidence, allocation rules, and a conservative scenario together. These figures illustrate the arithmetic; they are not benchmarks or forecasts. Add any later project costs incurred during follow-up.
You can also calculate a planning threshold. At CNY 10,000 contribution profit per order and a 50% GEO share, each order contributes CNY 5,000 toward the CNY 12,000 investment. Covering the cost requires at least three orders. At an assumed 10% conversion rate from qualified inquiry to order, that suggests roughly 30 qualified inquiries for planning purposes. Actual results will vary; an assumed conversion rate is not a sales commitment.
Allow for the sales cycle
Track leads by the month they entered the pipeline and follow each cohort through the normal sales cycle. Comparing this month’s spending with orders from older customers can distort the picture. Open proposals remain pipeline, not realized revenue.
Early measures can include cost per qualified inquiry and cost per sales-accepted opportunity, using the corresponding cohort’s costs. Historical conversion rates and contribution margins can inform forecasts, clearly labeled as estimates. With no qualified inquiries, report that directly; cost per inquiry is undefined, not zero.
Time saved on repetitive presales questions also matters. Measure the actual hours saved and report them separately. Freed capacity is an efficiency benefit; it becomes cash savings only when spending actually falls. Avoid counting the same benefit again in order profitability.
The six items to review each month
Cumulative investment: Cash spending, internal hours, and full cost.
Visibility changes: Citations and accuracy across the fixed question sample.
Qualified inquiries: Volume, source evidence, and cost per inquiry.
Sales progress: Accepted opportunities, proposals, wins, and loss reasons.
Attributed contribution profit and ROI: Tracked by lead cohort, with assumptions visible.
Next decision: Which activity to expand, change, or pause.
If citations rise but inquiries do not, examine whether the questions reflect purchase intent and whether the pages offer an appropriate next step. If inquiries rise but are poorly qualified, revisit positioning and content promises. If good opportunities remain open, follow them through the normal sales cycle.
Reduce or pause expansion when the agreed observation period and relevant sales cycle have passed, qualified demand remains weak, and costs exceed the agreed tolerance. Increase spending when opportunity quality and economics justify it.
Make the first investment useful
For a business with limited resources, a strong starting point is to turn information scattered across sales laptops, old brochures, and promotional images into website content that customers can read and verify.
That requires cooperation between IT, marketing, and the people who know the product. It also requires a manager who can define a manageable scope. A substantial new software or agency purchase may not be necessary to begin.
At your next project meeting, set one concrete objective: choose five pages, complete the information buyers need, verify access, and observe whether AI answers represent those pages accurately over time.
Share the pilot brief with your website owner and marketing lead. Identify what you can fix now, which facts need verification, and where specialist help would resolve a specific problem. That is a practical basis for deciding what to invest next.
