The third invoice is paid. The end-of-month report arrives: schema setup done, llms.txt live, six new pages written, every line on the technical checklist green. You close the report, pick up your phone and ask ChatGPT the most ordinary question in your sector — the one your customer asks before they ever come to you. There are three names in the answer. None of them is you. The second one is the firm you have been competing with in the same district for three years.
Two feelings arrive at once at this point: "the money was wasted" and "maybe I just need to wait a bit longer". Neither is a decision; both are feelings. To make a decision you first have to draw one distinction.
"No results" hides two completely different faults
When you say "no results" about a GEO engagement, one of two entirely different situations sits underneath it:
First: the work genuinely was not done. The items are written in the report but there is no trace of them on the site, or what was done was done wrong. Schema was added but it is faulty. Content was written but it answers a question nobody asks. Waiting changes nothing here; six months from now you will be in the same place.
Second: the work was done, but it is being measured in the wrong place. The model has started naming you on certain questions, just not on the one you tried. Or it does name you, but because it does not cite you as a source no clicks reach the site, and you read that as "nothing is happening".
These two do not have the same fix. The first is an execution problem, the second is a measurement and expectation problem. Before you change agencies, cut the budget or extend the contract, you need to know which one you are in. The whole point of this article is to let you draw that line.
A realistic timeline: implementation time and reflection time are not the same thing
The most common disappointment in GEO comes from one wrong assumption: "we made the change, so we should see the result."
In reality two separate clocks are running:
The implementation clock. Technical fixes — schema markup, llms.txt, robots.txt permissions, restructuring pages so they actually answer questions — are relatively fast work. The band cited in industry sources is a few weeks with a team that functions. This clock is under the agency's control and proof can be shown to you.
The reflection clock. Getting the change through to the language models is a separate process, and it is under neither your control nor the agency's. Some systems search live and pick a change up relatively quickly; others are tied to training and indexing cycles and are slower. The band cited in sources runs from a few weeks to a few months depending on the platform.
Lay the two clocks on top of each other and this is the picture:
| Stage | Band cited in sources | What it is realistic to expect |
|---|---|---|
| Technical fixes implemented | A few weeks | Visible, verifiable change on the site |
| Models reflecting it | A few weeks to a few months, by platform | Intermittent appearances on some questions |
| First concrete visibility shift | 4-8 weeks in most cases | Consistent mentions on specific question groups |
| Durable position | 6-12 months of disciplined work | Regularly cited on most of the core questions |
Do not read these bands as a commitment. They all shift with how crowded the competition is, how much authority your brand already has off your own site, and how often your sector gets asked about at all. A newly founded brand and a brand that has been in the press for ten years do not move on the same timeline.
But the practical conclusion from the table is clear: three months is too early to say "there is no signal" and too late to say "there is no evidence". Expecting a durable position in month three is not realistic. Seeing no intermediate signal, no measurement and no evidence at all in month three is a finding in its own right.
What you do not track, you assume is not happening
Take this possibility seriously: there are results and you cannot see them.
This is the most irritating difference between GEO and traditional SEO. In search results you have a ranking, you look, you see it. With language models the answer is not identical on every query; it changes with how you phrased the question, with your account history, with which version of which model you landed on. Three tries from your own phone is not a measurement; it is three anecdotes.
A real measurement setup contains at least this much: a fixed, written question list, those questions asked at regular intervals under consistent conditions, the output recorded, and traffic arriving at your site from AI interfaces tracked as its own segment. None of this is magic; it just has to have been set up. We walked through how to set it up step by step in measuring AI traffic and conversion in GA4.
If your three-month agency report shows no trace of that setup — no question list, no before/after comparison, only a list of "work completed" — what you have is not an absence of results, it is an absence of measurement. The two are very different things, and the second is easier to fix.
The eight-item diagnostic list
Each of the eight items below is a known fault that can quietly kill a GEO engagement. "Quietly" matters: none of them throws an error message, none of them lights up red in a report. Each item ends with a single check you can run yourself. You can finish all eight in one afternoon.
1. AI crawlers may be blocked by the site
For the models to read your site, the crawler bots have to get in. Some off-the-shelf themes and security plugins shut these bots out by default and nobody notices.
Your check: Type yoursite.com/robots.txt into a browser. In the text that opens, look for the names GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot and Google-Extended. If any of them has Disallow: / underneath it, that engine is not reading your site at all. For what this file should look like and how it relates to llms.txt, see what llms.txt is and how to manage AI crawlers.
2. Your CDN or security layer may be blocking bots silently
Even with a clean robots.txt, the security layer in front of your site can turn the bot away at the door. Services like Cloudflare have a "block AI bots" setting, and on some plans it arrives switched on. In that case nothing is wrong on your site; the bot never came in.
Your check: Go into your CDN or security panel, find the AI bot blocking setting and read its state. If the agency holds the panel access, ask for a screenshot of that setting. "We don't have anything like that" is not an adequate answer; see the screen.
3. Schema exists but is wrong, or inconsistent with the page
Structured data makes it easier for models to interpret your page correctly. But schema that is written incorrectly, or that contradicts what appears on the page, does no good. Schema with an out-of-date address, an un-updated service list, or a claim to a service the page never mentions is common.
Your check: Put the URL of your homepage and one service page into a structured data validator. If no errors come up, go one step further: compare the schema fields you see against what is visibly on the page. A single field that does not match weakens the signal. Schema markup and E-E-A-T goes into which fields matter.
4. Everything known about your brand rests on your own site alone
If a model learns everything it knows about you from a single source — your site — it uses that information cautiously. Verifiable information is information that appears the same way in more than one independent place.
Your check: Search your brand name in quotation marks and take your own site and social accounts out of the results. What is left? Trade directories, professional chamber records, local news, business profiles, independent review platforms. If nothing is left, the real gap is not technical; your brand does not exist off your own site.
5. Content warms up before it answers the question
Models producing an answer prefer text that answers directly. A paragraph that opens with "In today's rapidly digitalising world..." does not actually answer anything.
Your check: Open a service page on your own site and read only the first three sentences. Do those three sentences, on their own, answer the question in the page's heading? If not, that page has nothing quotable in it. Repeat the same test on your five highest-revenue pages.
6. The question list being tracked is not what your customers ask
In some reports, the question list showing "success" is made up of questions nobody actually asks. Appearing on questions that contain your brand name is normal and brings no new customers.
Your check: Ask the agency for the question list it tracks. Then write down the first ten questions that came to you by phone or WhatsApp in the last two weeks. Put the two lists side by side. If they do not overlap, you are measuring something that is not true. Fixing that list produces results faster than writing new content in most cases.
7. Being mentioned is being confused with being cited
These are different things. A mention is your name appearing inside the model's answer. A citation is the model presenting your page as the source of the answer, with the user able to click through. A mention is a reputation signal; a citation is a traffic source.
Your check: Ask, one by one, which of the two every "visibility" number in the report refers to. If the agency cannot separate them, the measurement method is not defined. For the other signs of whether an agency is genuinely doing GEO, see the checklist in is your current SEO agency actually doing GEO.
8. The page comes back empty without JavaScript
Sites whose content is produced entirely in the browser can serve a blank page to a simple crawler bot. You see a full page; the bot sees nothing.
Your check: Open the page, right-click, choose "view page source" and search the text that opens for a sentence from your page. If the sentence does not appear in the source code, your content is not coming from the server. This is an infrastructure problem, not a content problem, and it has to go to the technical team.
"Mentioned but no traffic" is not a failure
One of the most common outcomes during diagnosis is this: your brand appears in some answers, but there is no meaningful movement in analytics.
That does not mean the work failed. It means you are at a different stage. The model has started to recognise you but does not yet trust you enough to cite you; or the question closes with the answer itself and the user has no reason to click.
The right move at this stage is not to cut the budget but to change the target: turn mentions into citations. That usually takes three things — content blocks that answer directly and stand on their own; a verifiable brand trace off your own site; and original data or experience a model would find worth citing. The wrong move is to say "no traffic is coming" and reset everything, because if you reset you lose the partial recognition too.
The threshold for changing agencies: which three things are missing
Patience is not a virtue, it is a decision, and a decision needs a criterion. If at the end of the third month you have none of the following three things, changing agencies is a rational decision rather than an emotional one:
- A written, fixed definition of measurement. Which questions, on which platforms, at what frequency, and what counts as success? If that definition is not there at the start of the contract, an argument at the end of it is unavoidable.
- Verifiable evidence of the work done. Which page, on which date, what changed? Can a URL be produced, can a before/after be shown? "Optimisation was carried out" is not evidence.
- The ability to diagnose. When you put the eight items above to the agency, can it answer each one concretely, or does it retreat into general lines like "the algorithm keeps changing"? A team that cannot name the problem cannot fix it.
If all three are missing, change. If one or two are there, first propose running the diagnostic list in this article together — in most cases the problem is not bad faith but an undefined expectation. And if the measurement definition is not put in writing before the contract with a new agency either, you will spend the second three months in exactly the same fog.
An ordered plan for the next 90 days
What you do after the diagnosis matters, and so does the order you do it in. Content written before the blockage is cleared is wasted.
Weeks 1-2 — Diagnose. Work through all eight items and write the results on a single page. Base it on your own findings, not on the agency's report.
Weeks 3-4 — Clear the blockages. robots.txt, the CDN bot setting and server-side rendering problems get fixed first. No content work done before those are resolved can be measured.
Weeks 4-6 — Set up measurement. Write the fixed question list, define AI-sourced traffic as its own segment, take the first baseline reading. Without that baseline you will not be able to prove any improvement from here on.
Weeks 6-12 — Fix the content. Before producing new pages, rewrite your five most valuable existing pages so they answer directly. Update directories, profiles and records for off-site verification.
Week 12 — Measure again and compare. Ask the same question list under the same conditions. Compare the change against the baseline. This time what you have in your hand is data, not a feeling.
If you want to talk through your findings once you have run the diagnostic list, get in touch — let us name what is missing first, and the fix comes after that.
Frequently Asked Questions
When do the first results from GEO work show up?
The band cited in industry sources puts the first concrete visibility shift at four to eight weeks in most cases, while a durable position takes six to twelve months of disciplined work. These are not a commitment but an observed range, and they shift with how crowded the competition is, how much authority the brand already has off its own site, and how often the sector gets asked about. Expecting a durable position in month three is not realistic, but seeing no intermediate signal and no measurement at all in month three is a finding in its own right.
Should I change agencies if nothing has changed after three months?
Before you change, check whether you have three things: a written and fixed definition of measurement, verifiable evidence of the work done, and the ability to name the problem. If all three are missing, changing is a rational decision. If one or two are there, the problem is usually an undefined expectation rather than bad faith, and running the diagnostic list together produces faster results. If you change agencies without putting the measurement definition in writing again, you will spend the second three months in the same uncertainty.
Does not seeing my brand when I ask ChatGPT mean there are no results?
No. A language model's answer is not identical on every query; it changes with how you phrased the question, with your account history, and with which model version you landed on. Three tries from your own phone is not a measurement, it is three anecdotes. A real assessment requires a fixed question list asked at regular intervals under consistent conditions, the output recorded, and traffic from AI interfaces tracked as its own segment.
My brand appears in AI answers but no traffic reaches my site — is that a failure?
No, it is a different stage. A mention is your name appearing in the model's answer; a citation is the model presenting your page as the source with the user able to click through. A mention is a reputation signal and a citation is a traffic source, and the first usually comes before the second. The right move at this stage is not to cut the budget but to shift the target from mentions to citations; if you reset, you lose the partial recognition you have already earned.
How do I check that my site is open to AI crawlers?
Check the two layers separately. First type yoursite.com/robots.txt into your browser and look in the text for the names GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot and Google-Extended; if any of them has a Disallow directive underneath it, that engine is not reading your site. Then look at the state of the AI bot blocking setting in your CDN or security panel, because even with a clean robots.txt the security layer can turn the bot away at the door, and in that case nothing is wrong on the site yet the content is never read at all.