Chapter 17: GEO as a Continuous System

Last updated: November 1, 2026.

Picture a GEO project that’s finished. The prompt list is built, visibility has been measured, AI traffic is broken out in Analytics, the bots can get in, the pages are readable, and the brand details and third-party sources have been cleaned up. Run the same measurement three or four months later and you’ll find that part of that work has already gone stale.

AI answers change every day. Models get updated, search results move, competitors publish new pages, and your own prices and products change. That was true in classic SEO too, but in AI search things move faster and are harder to see from the outside.

So GEO has to be set up as a loop that keeps running, not as a project: Prompt → Measure → Diagnose → Fix → Measure again. The loop starts with an AI Search audit, and monthly, quarterly and yearly checks set its pace. At the end we’ll look at how the work gets split up inside an agency.

You’ll see a lot of references to earlier chapters. Prompt research was the subject of Chapter 6, the five visibility metrics of Chapter 7, and traffic and business results of Chapter 8. I won’t repeat them here. This chapter connects them.

1. Why a one-off project goes stale

Answers vary. In Chapter 3 we looked at the SparkToro and Gumshoe study: the chance of getting the same list of brands twice for the same question is under 1% (Search Engine Journal). A single measurement is a snapshot of one day. Ask the same question tomorrow and you may get a different list.

The ground keeps moving, too. According to Google’s Search Status Dashboard, the first ten months of 2026 alone brought one Discover update, two core updates (March and May) and four spam updates (March, June, August and September) (Google Search Status Dashboard). Because Google’s AI features rely on its ranking systems, any of these updates can affect your AI visibility as well. Then add the model updates from OpenAI, Google and everyone else.

And the business itself changes. A new product launches, a price changes, a branch closes, a service is dropped. AI learns these things from your site and from what’s written about you. Without a routine that keeps that information current, it goes on telling people the old version on your behalf.

AI visibility isn’t something you win once and put on the shelf, so part of the budget has to go to a recurring loop rather than to one-off work.

2. The five steps of the loop

1. Prompt: Decide which questions you want to show up for, and keep that list fixed.
2. Measure: Run those questions on each platform regularly, more than once, and record the results.
3. Diagnose: Where you don’t appear, or where you’re described wrongly, find the cause.
4. Fix: Depending on the cause, do the technical, content, brand or third-party work.
5. Measure again: Re-measure the same questions the same way and record the change.

There are two rules. Don’t skip steps; in my view the biggest risk is jumping from no measurement straight to fixes, or producing content before you’ve diagnosed anything. And close the loop. If you fix things and never measure again, you won’t know whether the work did anything.

3. Managing the prompt list

The loop starts with the prompt list. Chapter 6 covered how to build it: the questions your sales team hears, the customer’s constraints, persona clusters, and 30 to 50 commercial prompts to start.

For the loop, how the list is managed matters more than what’s on it. Keep the core list fixed; the prompts you measure at the start of the month should be the same ones you measure at the end. If you change the list every month, you can’t tell whether a change came from your work or from the list. Note every prompt you add or remove, with the date. A year from now, that note will save you a lot of guessing.

When a new product, a new service or a new question from the sales team comes up, put it in a separate test group. Measure new prompts there for a few cycles, then move them into the core list.

Next to each prompt, write down which product or service it relates to, which persona would ask it, and what it’s worth to the business. Diagnosis and prioritization will depend on this. For the filling-machine manufacturer, the core list might include prompts like “filling machine for 3,000 bottles a day of a foaming product,” “filling machine manufacturers with service in Turkey” and “liquid filling machine on a €25,000 budget.” Each one maps to a persona and a product group.

Taking a prompt off the list is a decision too. If a product has been discontinued, a service dropped, or a market is no longer a target, remove the related prompts from the core list but keep their past measurements in an archive. As prompts with no business value pile up, measurement gets more expensive and the report gets noisier.

4. Keep the measurement method the same

Chapter 7 covered measurement in detail: five metrics instead of a single score, running the same prompt several times, measuring each platform separately, and knowing the limits of the tools. Running it as a loop adds a few things.

The method has to stay the same every cycle. Same platforms, same language and location settings, same number of runs, and the same tool if possible. If the method changes, the results change, and you might mistake that for a real shift.

Keep the raw data. A summary like “35% visibility” isn’t enough. Save the answer itself, the brands mentioned, the pages cited and the date. Diagnosis runs on this data. Looking at which pages get cited is the easiest way to see what content is putting competitors ahead of you.

Use the platforms’ own data as well. Two things happened on this front in 2026. On June 3, 2026, Google added generative AI performance reports to Search Console and, as of August 31, 2026, opened them to all sites (Google Search Central). The report shows impressions in AI Overviews, AI Mode and the AI features in Discover, broken down by page, country, device and date. For now it only has impressions; there’s no click or query breakdown.

Microsoft opened the AI Performance report in Bing Webmaster Tools as a public preview on February 10, 2026 (Bing Webmaster Blog). It shows how many times, and with which pages, your site was cited in Copilot, in Bing’s AI summaries and with some partners, along with a sample of the queries used to reach your content (grounding queries).

These two reports let you put the platforms’ own numbers next to the estimates from third-party tools. Google also issues a warning here: be careful with tools that promise ranking success or claim to use Google’s “internal” metrics, because no third-party tool has access to Google’s internal ranking or AI systems (Google Search Central). The tools show direction; the platform data shows what happened. Read them together.

Measurement also has a business side. If visibility goes up and demand doesn’t, either you’re showing up for the wrong prompts or the people who reach your site can’t find what they came for. In the monthly measurement, put AI-referred sessions and conversions from Analytics, the change in branded searches, and the “How did you hear about us?” answers from your quote form next to the visibility data. Chapter 8 covered all of this. Visibility is an intermediate goal. What you’re after is visibility that turns into sales or demand.

Be practical about the number of runs. Running every prompt dozens of times on every platform burns through the budget fast. For most businesses it’s enough to run the core list a few times per platform each cycle and read the results at the level of prompt groups rather than individual prompts. The bigger the group, the more the swings in individual answers even out.

5. Diagnosis: from symptom to cause

Measurement shows you a symptom. You don’t appear, or you’re mentioned but not recommended, or you’re described wrongly. Diagnosis means finding what’s behind it.

This is the step that gets skipped most easily. When the result says “we’re not showing up,” it’s tempting to go straight to writing new content. But the cause might be a bot block or a page that never got indexed, and then the new content won’t show up either.

It helps to ask the questions in this order:

1. Can the page be read? Is it indexed, can the bots reach it, is the content stuck behind JavaScript? (Chapters 2 and 9)
2. Does the page show up for the right searches? AI pulls its sources from search; if your page isn’t there for the sub-queries in the query fan-out, it isn’t being read. (Chapter 3)
3. Does the page answer the question’s constraints? Are capacity, price range, region and service spelled out on the page? (Chapters 10 and 11)
4. Does AI recognize your brand correctly? What it has in memory may be old or incomplete. (Chapter 13)
5. Who is describing you? If the cited pages include competitors’ comparison pages, forums or directories and you’re not on them, the problem isn’t on your site. It’s off it. (Chapter 14)

The order matters, because if there’s a technical block, content work is wasted, and if the content is missing, brand work is wasted.

The table below collects the most common symptoms, their likely causes and where to look first.

6. Which fix comes first

Diagnosis usually turns up more than one task, and you can’t do them all at once. To order them, look at two things: how much is this prompt worth to the business, and how much effort does the fix take?

In the filling-machine example, “filling machine manufacturers with service in Turkey” is a question that leads straight to a sale. If you’re not showing up because the site never mentions service at all, the fix is valuable and easy, so it goes first. Not being cited for an informational question like “what is a filling machine?” is low value and often expensive to fix.

Fixes usually fall into four groups. Technical covers bot access, indexing, sitemaps, speed and JavaScript. Content means adding missing information to existing pages, writing decision-support content and updating outdated information. Brand means making company details consistent everywhere and correcting wrong information. Third-party means being listed correctly in industry publications, directories and comparison sites.

Freshness needs a separate word of caution. According to an Ahrefs study published in July 2025 that looked at about 17 million citations, the pages cited by ChatGPT, Perplexity, Gemini and Copilot are on average 25.7% newer than the pages in organic search results. The same study found AI Overviews to be the platform that cares least about freshness (Ahrefs).

Don’t read that and start changing the dates on your pages. Google’s helpful content guidance explicitly lists changing a page’s date to make it look fresh, when the content hasn’t substantially changed, as a warning sign (Google Search Central). If a price changed, a new model came out or a regulation changed, update the page and its date. If nothing changed, leave it alone. Changing the date without adding new, accurate information isn’t an update.

Also separate quick fixes from long-term work. Quick fixes are usually technical or about missing information: removing a bot block, getting a page indexed, adding service details to a product page, correcting an old price. They take a few days and their effect can show up in the next measurement. Long-term work sits on the brand and third-party side: getting into an industry publication, publishing an original data study, getting more customer reviews. Those can take months to show results.

Put some of each into every cycle. A plan that only does quick fixes stalls after a while because the easy work runs out. A plan that only does long-term work shows nothing in the first months and loses support inside the company.

7. Don’t rush the re-measurement

After the fix, measure the same prompts the same way. But it takes time for a change to show up in AI answers; the page has to be recrawled and reindexed. Even Google’s own documentation says that when you change the AI setting in Search Console, the effect usually takes a few days and can take longer for some content because of caching and propagation across systems (Google Search Console Help). Information that comes from the model’s memory may not change until the next model update.

So don’t re-measure a week later. Do it in the next regular measurement cycle. When you read the results, look at the change across a group, not in a single prompt. Visibility for one prompt going from 20% to 40% may just be noise; a similar increase across most of the ten prompts in the same group means more. If there was a Google update, a model update or a big competitor campaign during the measurement period, note that too.

The most useful record is the change log: which page, what date, what changed. A simple table is enough, with columns for date, page, change made, expected effect and related prompts. Without it, months later you won’t be able to tie any increase or drop to a cause.

Accept from the start that you won’t be able to explain every result. AI answers depend on a lot of factors. The distinction from Chapter 8 between observed, estimated and modeled impact applies here as well. Keeping what you know separate from what you’re estimating in your reports is the best way to protect trust over the long run.

8. One full cycle

Let’s follow the loop through a hypothetical example, using the filling-machine manufacturer from earlier chapters. The numbers below are purely illustrative.

The core list has 40 prompts: 12 a purchasing manager might ask, 15 a production manager might ask, 8 from the maintenance team and 5 from finance. Each prompt is run five times in ChatGPT, Gemini, AI Mode and Copilot. The brand is mentioned often for production manager questions but hardly ever shows up for purchasing manager questions. In Copilot it doesn’t appear in any group.

Diagnosis turns up two separate problems. The purchasing manager’s questions about delivery times, service network and spare parts have no answers on the site; that information exists only in the sales team’s heads. And a look at Bing Webmaster Tools shows the site’s main product pages aren’t indexed in Bing.

The Bing indexing problem gets fixed first, because it’s easy and it affects all Copilot visibility. Then there’s a one-hour meeting with the sales team, and the service network, spare-parts lead times and delivery times get written onto the product pages and onto a separate service page. Every change goes into the change log with its date.

In the next measurement cycle, the brand starts appearing in Copilot for some prompts. Visibility for purchasing manager questions goes up but still trails the competitors. The cited pages show that the competitors are listed in an industry directory and this company isn’t. That becomes the diagnosis for the next cycle.

The loop never says “done.” Each cycle produces the task list for the next one.

9. Where to start: the AI Search audit

Before setting up the loop, it helps to do one thorough audit. Call it an AI Search audit. The goal is to take a snapshot of where the site and the brand stand today and produce the first task list for the loop. The audit can be organized under five headings:

• Technical access. Are the important pages indexed in Google and Bing? Are AI bots blocked in robots.txt, in a security layer like Cloudflare, or at the server? Can the content be read without JavaScript? Is the Search generative AI control in Search Console set to “include”? It’s on by default for every site, but it’s worth checking that nobody turned it off (Google Search Console Help).
• Content coverage. Is there a page on the site that answers each question on the core list? Do the product and service pages answer the user’s constraints?
• Brand and data consistency. Are the company name, address, phone number, products, prices and service areas the same on the site, in the Business Profile and in external sources?
• Third-party presence. How is the brand described on the industry sites, directories and forums that AI cites?
• Measurement setup. Is AI-referred traffic broken out in Analytics? Are the AI reports in Search Console and Bing Webmaster Tools being tracked? Does the quote form ask “How did you hear about us?”

The output of the audit should be a prioritized task list, not a long report. In my view, a good audit report is one where the reader knows what to do the next morning.

A reasonable order for the first month looks like this. In week one, remove the technical blocks, since everything else depends on them. In week two, finalize the core prompt list and run the first measurement; that becomes your baseline. In weeks three and four, make the diagnosis and write the plan for the first quarter. If you start without a baseline, three months later you won’t be able to answer “what changed?”

Who does the audit matters too. The site owner, or a team that works on the site every day, can get used to its gaps. If possible, have a fresh pair of eyes do the yearly audit: another team inside the agency, another department inside the company, or an outside specialist.

10. Monthly, quarterly and yearly checks

Doing everything every month is expensive and unnecessary. Here’s the schedule I suggest:

The first row is the most underrated check. As I mentioned in Chapter 2, one of our clients had their site go down for about an hour, and since then we at İmza monitor uptime for every client’s site around the clock. If the site is down when a search engine or AI bot comes by, there’s nothing to read at that moment.

If the monthly measurement shows a big drop, don’t jump straight to fixes. Repeat the same measurement a few days later first; with answers this volatile, a single measurement can mislead you. If the drop holds in the second measurement, move on to diagnosis.

Factor seasonality into the schedule. For an e-commerce site, the November sales period; for a tourism business, the months before the season; for a B2B manufacturer, the weeks before major trade fairs may call for more frequent measurement and faster fixes. In those periods, checking that price, stock and campaign information shows up correctly in AI answers is worth more than the measurement in an ordinary month.

11. Splitting the work in an agency

This loop isn’t one person’s job. In an agency or an in-house team, the work can be divided roughly like this:

• The SEO specialist owns the loop. They manage the prompt list, run or commission the measurement, make the diagnosis and set priorities.
• The technical team handles fixes to bot access, indexing, site speed, sitemaps and structured data.
• The content team adds missing information to pages, prepares decision-support content and updates old content.
• On the client side, the sales team supplies the questions and answers and confirms that product and price information is accurate. Without them, good content doesn’t happen.
• PR and communications work on getting listed correctly in third-party sources.

On the agency side, the part of the job that takes the most effort is reporting. A monthly report only needs to answer four questions:

1. What changed? What happened with visibility, citations, AI-referred traffic and conversions?
2. Why did it change? Which causes do we know, and which are we estimating? Write them separately.
3. What did we do? The work done this month, taken from the change log.
4. What’s next? Next month’s priorities and what we need from the client.

The language of the report affects the outcome too. As I said in the first chapter, “we’ll get you to number one in ChatGPT” doesn’t fit how this works. The report should talk about rates, not positions, for example: “This month you were recommended in 14 of the 40 prompts on the core list, up from 9 last month.” What you can promise a client is a regular, transparent process, not a result.

Agree on what you need from the client up front. There’s work in this loop the agency can’t do on its own: confirming price and product information, freeing up the sales team’s time, making room in the development team’s schedule. A loop set up without talking about these things gets stuck at the diagnosis step. A short monthly call often gets more done than a long report.

Be realistic about cost as well. The most expensive part of this loop isn’t measurement, it’s fixing. Keep the measurement budget limited and put most of the resources into the work the diagnosis uncovers. A project that measures everything well and fixes nothing is just an expensive monitoring system.

12. Start small and keep it turning

You don’t have to set up everything in this chapter in the first month. A core list of 30 to 50 prompts, a monthly measurement, a simple change log and uptime monitoring are enough to start. After a few cycles you’ll see for yourself which checks tell you something and which just pad the report.

This chapter wraps up Part 5 of the book. In the last two chapters we turn things around and look at how to use AI in SEO work itself.

Checklist

Sources

• Google Search Central Blog: Introducing Search Generative AI performance reports in Search Console (June 3, 2026; available to all sites as of August 31, 2026)
• Bing Webmaster Blog: Introducing AI Performance in Bing Webmaster Tools (Public Preview) (February 10, 2026)
• Google Search Central: Optimizing your website for generative AI features on Google Search (last updated July 10, 2026)
• Google Search Console Help: Search generative AI control (available to all sites as of August 31, 2026)
• Google Search Status Dashboard: Ranking incidents history (checked November 1, 2026)
• Google Search Central: Creating helpful, reliable, people-first content (last updated October 5, 2026)
• Ahrefs: AI Assistants Prefer to Cite “Fresher” Content (17 Million Citations Analyzed) (July 28, 2025)
• Search Engine Journal: AI recommendations change with nearly every query (SparkToro) (January 30, 2026)

Previous: Chapter 16, AI Search by Industry · Table of contents · Next: Chapter 18, AI in SEO Work