Guides
SEO rank tracking vs GEO tracking: what changes in the data
A SERP row carries an integer position, an AI answer carries none. Two runs of one prompt repeat about 40% of their cited domains, so AI visibility needs samples.
By Mathias C.SEO consultant, and more recently GEO, at MentionBox
Rank tracking and GEO tracking fill two different tables. A rank row is a keyword, a location, a device and one integer: DataForSEO's Google Organic response puts rank_group and rank_absolute on every item, and one scrape settles the value. A GEO row is a prompt, an engine, a locale and a run, and it carries no position at all, because the answer payloads published by OpenAI, Anthropic and Perplexity return a URL, a title and a date per source and nothing resembling a rank. Every figure in an AI visibility report is therefore counted afterwards, over a stored set of answers, and the set has to be large: cloro's study of August 14, 2026 measured a mean overlap of 0.40 between the cited domains of two runs of the same prompt minutes apart, and puts about 300 answers per period at roughly ±5 percentage points.
This page is for whoever already runs a rank tracker and now has to produce an AI visibility report beside it, and for whoever writes the code behind both tables. It goes through the row key, the missing position field, the sampling, the query column that disappears, the metrics the vendor computes, the free first-party reports on each side, the published cost per 1,000 rows, and the habits that break on the way across.
The row key changes, and it gets longer
| Dimension | Rank tracking row | AI visibility row |
|---|---|---|
| What you pick | a keyword, with a measured monthly volume | a prompt, with no measured volume |
| Where | location_code and language_code | country and language of the run |
| On what | device, desktop or mobile | engine and model |
| When | the date of the check | the timestamp of each run |
| Repeats | one check settles the value | several runs per period, averaged |
| The value stored | an integer position among the results | the answer text, its sources, the brands named in it |
Row dimensions from DataForSEO's Google Organic live advanced reference and from the engine API documentation linked below, read on November 6, 2026.
Two of those lines decide how big the dataset gets. On the SEO side the row count multiplies with locations and devices, and the check is daily by habit, so 500 keywords on one location and one device is 15,000 rows a month. On the AI side it multiplies with engines and with runs, and the runs are not a refinement: they are what makes a percentage mean anything. Four engines and six runs turn a 50-prompt list into 1,200 answers a month, which is 300 per engine.
An AI answer has no position field
The rank field is explicit in a SERP payload. DataForSEO defines rank_group as the "position within a group of elements with identical type values" and rank_absolute as the "absolute position among all the elements in SERP", and bills "per each SERP containing up to 10 results":
{
"keyword": "best crm for small business",
"location_code": 2840,
"language_code": "en",
"items": [
{
"type": "organic",
"rank_group": 1,
"rank_absolute": 3,
"domain": "example.com",
"url": "https://example.com/crm",
"title": "CRM for small teams"
}
]
}
The equivalent object on the AI side is a block of text with sources attached. Anthropic's web search tool returns a server_tool_use block holding the query the model wrote, then web_search_result entries carrying url, title, page_age and encrypted_content, then citations with up to 150 characters of cited_text:
{
"type": "web_search_tool_result",
"content": [
{
"type": "web_search_result",
"url": "https://example.com/crm",
"title": "CRM for small teams",
"page_age": "October 2, 2026"
}
]
}
The other engine APIs are built the same way. OpenAI's web search tool attaches url_citation annotations that contain "the URL, title and location of the cited source", and a sources field that "returns the complete list of URLs the model consulted when forming its response", with the warning that "the number of sources is often greater than the number of citations". Perplexity's API returns search results with a title, a URL and a date. None of the three exposes an ordering you can report as a rank.
What replaces the rank is a set of yes-or-no facts about one answer: was your brand named, where in the text, which competitors were named with it, was one of your URLs cited, what tone was used. Turning those into a percentage is the vendor's job or yours, and the formulas differ by vendor, which our guide on share of voice, visibility and position works through field by field.
One check settles a keyword, one run settles nothing
A keyword check gives you a value you can put in a cell. A prompt run gives you one draw from a distribution. SparkToro's published research ran 12 prompts through ChatGPT, Claude and Google's AI with 600 volunteers over November and December 2025, for 2,961 combined runs, and reports "a <1 in 100 chance that ChatGPT or Google's AI, if asked 100X, will give you the same list of brands in any two responses"; on ordering, "it's more like 1 in 1,000 runs before you'd see two lists in the same order". The same study found the brands at the top of the list stable across differently worded human prompts, appearing 55% to 77% of the time, so the instability sits in the exact list rather than in the whole picture.
cloro's figures put numbers on the sample that follows from it. Two runs of the same prompt on the same model, minutes apart, repeated on average 0.40 of their cited domains, with per-prompt overlap from 0.06 to 1.00. Seventy-two answers carry a 95% confidence interval of about ±10 percentage points, about 300 bring that to ±5, and about 294 answers in each period are needed before a move from 20% to 30% is a result rather than a reading.
The query column disappears
A rank tracker and Search Console both hand you the string a person typed. In AI search the engine writes its own searches, which Google calls query fan-out, defined in its documentation on AI features as "issuing multiple related searches across subtopics and data sources". The Anthropic payload above makes the same point concretely: the input.query in the server_tool_use block is the model's phrasing, not the user's prompt. Our explainer on query fan-out covers how those sub-searches decide which pages end up cited.
Two things follow for the data. The prompt list is a sample you chose, so its composition is now part of your methodology rather than an observation of demand, and adding prompts moves every percentage computed over it. And prompt volume does not exist as a measured count: keyword volume comes from Google's advertising data, which DataForSEO's search volume endpoint takes from the Google Ads API for up to 1,000 keywords per request, while Ahrefs builds AI impressions by "summing up search volumes (based on Google) of prompts where your brand appears as an AI Answer", and DataForSEO's AI Keyword Data API returns volumes "reflecting their estimated usage in AI tools". Both are models, and both are honest about it.
The metrics come from the measurement vendor
Nothing in an engine response is a metric. Share of voice, visibility, position in the answer, sentiment and citation rate are counts computed over stored answers, and the denominator is a decision rather than a fact. A GEO API that returns them makes the decision for you and publishes it: API for GEO, which our reviews score 9.0 out of 10 and which we price at $13 per 1,000 answers down to $7 at its largest tier (checked on October 23, 2026), documents on its AI Visibility API page visibility as the "percentage of in-scope responses mentioning the entity", shareOfVoice as "entity mentions ÷ all entity mentions in scope" and avgPosition as the "mean rank of the first mention inside the answer". MentionBox, which publishes this site, has a commercial link with API for GEO.
The structural difference with a rank tracker is that the filters sent with the request define the denominator, so a date range, an engine or a prompt tag changes all the percentages at once. A rank report has no equivalent: filtering a rank table to one device does not redefine position 3.
- Store two counters per engine and per period, answers naming you and answers collected, and keep the raw answers. A stored 34.0% cannot be re-aggregated across engines without averaging averages, and a change in how you count a mention then costs a recount instead of a fresh month of collection.
- Freeze the prompt list for the length of the comparison. Changing it mid-quarter moves every percentage with nothing having changed in the answers.
What Google and Microsoft publish for free on each side
| Free first-party source | Covers | Metrics | Dimensions |
|---|---|---|---|
| Search Console Search Analytics API | Google Search, with AI feature traffic folded into the Web type | clicks, impressions, CTR, average position | query, page, country, device, searchAppearance, date, hour, up to 25,000 rows |
| Search Console Generative AI performance report | AI Overviews, AI Mode, generative AI in Discover | impressions only, no clicks, no CTR, no position | page, country, date, device in the Search report |
| Bing Webmaster Tools AI Performance | Microsoft Copilot, Bing and select partner AI experiences | citations, citation share per grounding query, no clicks | grounding query, intent, topic |
Report contents from Google's and Microsoft's own documentation, read on November 6, 2026. The generative AI report finished rolling out to all websites worldwide on August 31, 2026; Microsoft's AI Performance report launched in February 2026 and gained intents, topics and citation share on June 16, 2026.
Three details are worth reading twice. Google states that AI Overviews "are only shown when our systems determine that it is additive to classic Search, and as such, often don't trigger", so an empty cell on the AI side can mean the surface did not appear rather than that you lost. Microsoft defines citation share as "the percentage of citations attributed to your site out of all citations shown across all sites for that same grounding query", which is a share of citation slots on a query Copilot generated, not a share of brand mentions in an answer. And neither report carries clicks, which leaves the traffic question to the ordinary Search Console report, where AI clicks sit mixed in with everything else.
On that traffic question the published measurement to quote is Pew Research's study of July 22, 2025: across 68,879 Google searches by 900 US adults in March 2025, of which 12,593 produced an AI summary, users "clicked on a traditional search result link in 8% of all visits" when a summary was present, against 15% when it was not, and clicked a link inside the summary in 1% of visits.
What a row costs on each side
| Dataset | Published rate | Per 1,000 rows |
|---|---|---|
| Google SERP with positions, DataForSEO standard queue | $0.0006 per SERP of 10 results | $0.60 |
| Google AI Mode answer, DataForSEO AI Mode SERP API | from $0.0012 per task | $1.20 |
| Claude answer through the official API | $10 per 1,000 web searches, plus token costs | $10.00 and up |
| AI answer with the metrics computed, API for GEO | $13 per 1,000, $7 at the largest tier | $13.00 |
Rates from each vendor's pricing page and documentation. API for GEO's prices are from our price check of October 23, 2026; Anthropic bills per search rather than per answer, and its documentation puts single-fact queries at one to three searches and comparative research at ten or more.
Put a month through those rates and the shape of the two budgets shows up:
- Rank tracking. 500 keywords, one Google check a day, one location, one device: 15,000 rows at $0.0006 is $9.00 a month, and every row carries a position you can chart on its own.
- Raw AI answers. 50 prompts on four engines, six runs each: 1,200 answers at $1.20 per 1,000 is $1.44 a month, with the brand matching, the ranking of mentions and the arithmetic still to write.
- AI answers with the metrics in the response. The same 1,200 answers at $13 per 1,000 is $15.60 a month, and the 300 answers per engine land at about ±5 points by cloro's figures.
The per-row price is not what decides an AI budget at this scale, the run count is. Our cost breakdown for 100 prompts on five engines runs the same arithmetic across more APIs, and the ranking of AI visibility APIs scores them on coverage, data returned, price and integration.
Six reporting habits that break on the way across
- Charting a single prompt like a single keyword. One prompt's result is a draw; the reportable unit is a rate over a set of prompts and runs.
- Reading position as a rank. In an AI report it is the mean rank of your first mention among brands named in the text, averaged only over answers that name you, so it can improve while you disappear.
- Comparing two vendors' visibility percentages. Denominators, prompt lists, engines, countries and run counts all differ, so the two numbers describe two populations.
- Treating a missing AI answer as a loss. Per Google, AI Overviews often do not trigger, and a surface that did not appear is a different event from an answer that ignored you.
- Using keyword volume as prompt weight without saying so. It is a documented modelling choice at Ahrefs and DataForSEO, and it makes high-volume prompts count for more than niche ones.
- Pricing the AI dataset per answer. Multiply by engines and by runs first, then decide how many prompts the budget allows.
Our view
In our view the two datasets answer two questions, and the mistake is to put them in one chart. Rank tracking answers "where does this URL sit for this keyword today", with a settled value per row and a free first-party check in Search Console. AI visibility answers "how often, and in what company, does a brand come up when this kind of question is asked", which is a survey result and carries a margin of error whether or not the tool prints one.
Frequently asked questions
What is the difference between SEO rank tracking and AI visibility tracking?
Rank tracking records the position of a URL in an ordered list of results for a keyword, and the position is an integer returned by the scrape: DataForSEO's Google Organic response defines `rank_group` as the "position within a group of elements with identical type values" and `rank_absolute` as the "absolute position among all the elements in SERP". AI visibility tracking records whether your brand or your page appears inside a generated answer to a prompt, and the engine APIs return no rank at all, only the answer text and the sources behind it. One is read off a single row, the other is counted over a stored set of answers by the measurement vendor.
Can I track my position in ChatGPT the way I track a keyword ranking?
No, because there is no ordered result list to read a position from. What vendors call position in an AI report is the rank of your brand's first mention among the brands named inside the answer, averaged over the answers that name you, so it is a property of the text rather than of a results page. It also improves when your visibility falls, since the average is taken only over the answers where you appear, which is why it belongs next to the appearance count.
Does Google Search Console show AI Overviews and AI Mode data?
Partly. Clicks and impressions from AI features are folded into ordinary Search traffic: Google states that pages shown in AI features are included in the overall search traffic in Search Console, under the Web search type, and the Search Analytics API exposes no AI dimension among `country`, `device`, `page`, `query`, `searchAppearance`, `date` and `hour`. The separate Generative AI performance report, rolled out to all websites worldwide on August 31, 2026, breaks out impressions from AI Overviews, AI Mode and generative AI features in Discover by page, country and date, and reports no clicks, no queries and no position.
How many AI answers do I need before a visibility percentage means anything?
Several hundred per period and per engine. cloro's published study of August 14, 2026 puts 72 answers at a 95% confidence interval of about ±10 percentage points, about 300 answers at roughly ±5 points, and about 294 answers in each period to support a claimed move from 20% to 30%. A list of 50 prompts run six times a month on one engine reaches 300 answers, which is a quarterly comparison rather than a weekly one.
Is AI visibility tracking more expensive than rank tracking?
Per row, yes, and the run count widens the gap. A Google SERP with its positions starts at $0.0006 for 10 results on DataForSEO's standard queue, a Google AI Mode answer starts at $0.0012 per task, Anthropic's web search tool is billed at $10 per 1,000 searches plus token costs, and an answer with share of voice, position and sentiment already computed runs $13 per 1,000 on API for GEO, down to $7 at its largest tier, checked on October 23, 2026. The multiplier that decides a budget is the number of runs: 500 keywords checked once a day is 15,000 settled rows, while 50 prompts on four engines need six runs each to reach 300 answers per engine.
Do keyword search volumes exist for AI prompts?
Not as measured counts. Keyword volumes come from Google's own advertising data, which DataForSEO's search volume endpoint takes from the Google Ads API, while prompt volumes are modelled: Ahrefs builds its AI impressions by "summing up search volumes (based on Google) of prompts where your brand appears as an AI Answer", and DataForSEO's AI Keyword Data API returns volumes "reflecting their estimated usage in AI tools". Treat a prompt list as a sample you chose rather than as a map of demand.