If you already track which sources AI engines cite, the next question is usually how finely to look at them. You can group citations by website, or you can keep every cited page separate. That choice decides what your reports can answer, and it is easier to make once you see what each view leaves out.
Here is the short version. URL level citation tracking records which specific pages an AI answer cites. Domain citation tracking groups those same page citations under the website they came from. The domain view tells you which sites show up as sources. The URL view tells you which pages are behind that, and which tracked prompts cite them. These are two ways of adding up the same observations, so they are not really two competing methods, and you can often have both from one set of data.
The table below puts them side by side before we go through each row.
| Decision criterion | Domain-level tracking | URL-level tracking |
|---|---|---|
| What gets grouped | A domain bucket | One cited page |
| Main question | Which websites are cited? | Which pages are cited? |
| What it can hide | Which pages contribute, concentration, pages swapping places | The site's overall position, unless page records are rolled up |
| Change over time | Combined movement of the whole domain | Which pages gained, lost, or replaced each other |
| Competitor view | Which rival sites get cited often | Which specific rival pages get cited for your prompts |
| Main risk | Reading a site-wide total as broad content success | Reading a citation as proof of quality or business impact |
| Best fit | Decisions that stop at the website level | Decisions that need a specific content asset |
This article assumes you know what a citation is. If you want that background first, what AI citation tracking is covers it, and we start here from the difference in granularity. We also won't compare vendors' reports here, and we won't get into why a particular page does or doesn't get cited. Finding out which page was cited is attribution. Investigating why it performed that way is a separate job, usually called page level citation diagnostics, and it needs its own evidence.
Domain Citation Tracking and Source Visibility
Domain citation tracking answers a simple question: which websites do AI engines point to when they answer the prompts you care about? Every cited page gets assigned to its domain, and the domain gets the total. You end up with a short list of sites, often with a share next to each one.
That is a useful view when the decision is about websites. If a leader asks which publishers and competitors dominate your category's answers, a domain list answers it in one glance. It also makes a clean trend line. You can watch a site's total move up or down month to month without scrolling through hundreds of rows.
It works well for getting a source landscape, for reporting to people who care about websites more than pages, and for spotting which third-party domains deserve outreach, which is easier when you also find the sources AI engines cite in your industry.
The catch is that a domain total is a sum, and a sum can't tell you what went into it. The rest of this section covers the main things it hides.
Concentration
The same domain total can come from one page that gets cited again and again, or from a wide spread of pages that each get cited now and then. Those are different situations. If you're a marketing lead deciding where to spend editorial time, you'd want to know which one you're in.
A heavily cited page isn't a problem in itself. A good reference page can earn its place. What matters is whether your team knows where the citations are concentrated, because a domain total can't show you.
One page replacing another
A steady domain total doesn't mean the same pages stay visible. Here is a made-up example to show the idea. In one reporting window, Guide A receives all 20 of the citations your domain earns. In the next matched window, Comparison B receives all 20. The domain total is the same, and the cited content is completely different.
It only shows how information gets lost when you add things up, and it doesn't tell you what caused the change.
Content type and prompt differences
A domain total can't say whether the citations land on documentation, research, guides, product pages, or comparison pages, and a blended total can hide that different engines cite different pages for different questions. URL-level data helps only if each record keeps its prompt, engine, and collection window. With those attached, you can classify the cited pages yourself and see which pages are cited for which monitored questions.
Your pages and other people's pages
When a review site or publisher writes about your brand and an engine cites that page, the citation belongs to the third-party domain, even though the page is about you. It helps to keep citations to your own pages, citations to third-party pages about you, and brand mentions in the answer text apart in your reporting, and citation tracking tools for content teams explain the mention and citation split in more depth. One brand can also own several domains, so brand-level reporting needs an explicit ownership map.
URL-Level Citation Tracking and Page Attribution
URL level citation tracking keeps each cited page as its own row. For every citation you can see the page, how often it was cited, and which tracked prompts cited it. A domain total can then be rebuilt from those rows whenever you want one, which is why the finer record is the safer one to keep. The reverse doesn't work, because a domain total can't give you back the page distribution that was thrown away when it was added up.
This is the view you want when the decision is about content. Here is how the questions you might ask map to the evidence and a sensible next step.
| Your question | Useful evidence | A reasonable next step |
|---|---|---|
| Which of our content contributes most? | Page citation counts and each page's share of the domain | Pick the assets that deserve editorial attention and keep monitoring them |
| Is our visibility broad or concentrated? | Distinct cited-page count and the spread across pages | Decide whether the concentration is fine for your strategy |
| Did our cited content change? | Page records over matched windows | Look into which assets replaced others in the observed citations |
| Which questions lead to which pages? | Page-to-prompt relationships | Connect your editorial backlog to the buyer questions you monitor |
There are limits worth saying out loud. These observations help you prioritize and investigate, but they don't establish why an engine picked a page, they don't prove an optimization worked, and they don't promise that a similar page will earn citations. A page record also points to an asset. It doesn't necessarily point to the exact passage that shaped the answer, so don't describe it as passage-level attribution unless the data really contains that.
If you want to go from this view to action on the competitor side, tools that find AI citation opportunities usually work from page-level records.
Counts, Rates, and Denominators
Before you compare either view, it helps to be clear about what is being counted. The big difference is often not domain versus URL at all. It is what gets counted inside each grouping. The definitions below are for this article. Products don't all use the same names, denominators, or duplicate handling, so check how any report you use defines them.
| Metric | Definition | What it tells you |
|---|---|---|
| Citation occurrences | Qualifying citation references assigned to a domain or page | How often citation events happen under a stated repeat-counting rule |
| Response-level citation rate | Analyzed responses with at least one citation to the source, divided by all analyzed responses, times 100 | The percentage of analyzed responses that cite the source |
| Domain source share | The domain's citation occurrences divided by all qualifying citation occurrences in the dataset | The domain's portion of the measured citation pool |
| Page contribution within a domain | The page's citation occurrences divided by its domain's citation occurrences | How much that page adds to the domain's citations |
| Distinct cited-page count | Different pages cited within the domain | How many separate assets are cited, subject to how pages are identified |
A few counting rules are worth keeping in mind:
- An occurrence total and a response-level rate are different things, and a source share and a citation rate use different denominators.
- Several domains can appear in one response, so their response-level rates can add up to more than 100%.
- Page presence rates usually can't be added up to get domain presence, because pages can show up together in the same response.
- If a denominator is zero, report the rate as not available instead of quietly showing a zero.
One more point on repeats. If a single answer cites two different pages from the same domain, one common convention counts that as two domain citation occurrences, and another counts it as one response where the domain appeared. Both can be valid, and they answer different questions, so it helps to know which one a report uses.
A Worked Example With One Dataset and Two Views
Everything in this example is made up. It is not DeepSmith customer data and it is not an industry benchmark. Each listed reference counts as one citation occurrence.
| Analyzed response | Cited sources |
|---|---|
| Response 1 | Owned Guide A, Owned Comparison B |
| Response 2 | Owned Guide A |
| Response 3 | Another publisher's Research C |
| Response 4 | Owned Guide A, Owned Documentation D |
The domain view
For the domain you own, the domain view gives you these numbers:
- Five citation occurrences.
- Citations in three of the four responses, which is a 75% response-level citation rate.
- Five of the six citation occurrences in the whole example, which is an 83.33% source share.
The 75% and the 83.33% are both correct. They differ because the denominators differ: one divides by responses, and the other divides by all citations.
The URL view
| Owned page | Citation occurrences | Responses citing the page | Response-level citation rate | Share of owned-domain citations |
|---|---|---|---|---|
| Guide A | 3 | 3 | 75% | 60% |
| Comparison B | 1 | 1 | 25% | 20% |
| Documentation D | 1 | 1 | 25% | 20% |
| Total | 5 | Not additive | Not additive | 100% |
Now you can see what the domain view hides. Guide A supplies 60% of the owned domain's citation occurrences, and the domain has three distinct cited pages, not five. The three page rates add up to 125%, but the domain appears in only 75% of responses. That's because Responses 1 and 4 each cite more than one owned page. Domain presence has to count those responses once, which means taking the union of them, and a simple sum can't do that.
So a domain total tells you how much citation activity a site received, and the page distribution shows where that activity lives.
Domain and URL Identity Rules
Both views depend on decisions about identity that are easy to skip. This section is about measurement, not about technical page fixes.
What counts as a domain
A hostname isn't always the same grouping as a registrable domain. A registrable domain is often described as eTLD+1, the registrable name directly below an effective top-level domain. Multi-part suffixes matter here, and taking the last two dot-separated labels doesn't work everywhere. MDN's registrable domain glossary entry gives effective top-level domain examples such as .com, .org, and .ac.uk.
Before you trust a domain total, find out whether it combines the www version, subdomains such as blog and docs, regional domains, and separate properties on shared hosting. These are reporting choices. You can't infer them from the word "domain" alone, and a registrable-domain total isn't automatically a total for the whole company or brand.
What counts as a page
A URL has parts: a scheme, a hostname, a path, a query, and a fragment. Different strings can sometimes point to the same content, and strings that look almost the same can point to different content. That makes page identity a decision too.
| What happens | The risk | A sensible way to handle it |
|---|---|---|
| Tracking parameters create variants | One page is split across several rows | Keep the cited string and note which tracking parameters you normalize |
| A query parameter changes the content | Different pages get merged | Don't strip every query parameter automatically |
| An old address redirects | History is split between old and new addresses | Keep the originally cited address and the resolved destination as separate fields |
| A canonical annotation points to another version | Reports may merge variants differently | Say whether you use canonical mapping, and keep the original evidence |
These are suggestions for keeping an analysis transparent, and they don't describe how any particular product, DeepSmith included, normalizes URLs. Google describes redirects and canonical annotations as strong signals in its guidance on consolidating duplicate URLs, but that guidance is about how Google Search handles similar pages. It doesn't mean every AI engine or citation tool maps identity the same way.
What Both Views Can and Can't Tell You
A few limits apply no matter which granularity you pick.
Cited sources and consulted sources
OpenAI's web search documentation separates the inline citations in an answer from the full list of sources the model consulted, and the consulted list can be longer. So a page in the consulted list isn't necessarily a displayed citation, and a domain total built from consulted pages measures something different from one built from answer citations. That is documented for that API, so don't assume the same fields exist in every consumer AI interface.
Citations, traffic, and endorsement
A citation observation doesn't show that anyone clicked, visited, or bought anything, and a sample of tracked answers isn't a count of everything real people saw. It also doesn't prove the content is correct or that the engine rates the whole site highly, so be careful before reading "often cited" as "most trusted."
Sampling and matched windows
What you record describes a set of collected responses for specific prompts, engines, and dates. It isn't a census of every AI answer. Repeated answers can also change on their own. One preprint that repeated the same questions on Perplexity Search, OpenAI SearchGPT, and Google Gemini found that the overlap between cited domain sets from run to run was modest, with median overlap ranges of 0.29 to 0.31 for Gemini, 0.33 to 0.40 for SearchGPT, and 0.50 for Perplexity. The paper on uncertainty in AI visibility measured domain-set overlap, not URL overlap, and it covered three consumer-product topics with queries generated by ChatGPT. Treat it as a pointer to variability and not as a benchmark for your market, and don't mistake a stable-looking domain summary for a full account of what happened at the page level.
For comparisons to mean something, keep the prompt set, the engines, the collection frequency and window, the repeat-counting rules, and the domain and URL grouping steady, or say so when they change. Adding prompts or collecting more often can raise raw citation totals without improving how often you're present in responses. A response that was collected and had no qualifying citation also isn't the same as a failed collection, so don't count missing data as a confirmed absence.
It's also safer to say "not cited in the monitored sample" than "AI never cites this page." There's no universal minimum sample size or cadence that fits every program, so be wary of any rule that claims to give you one.
How DeepSmith Handles Both Granularities
DeepSmith is one platform for AI search analytics and content production, and it treats domain and URL as two views of the same citation data. You don't have to pick one and give up the other. Its source reporting can be shown at the domain level or the URL level, and sources are tagged as owned, competitor, or other, so your own pages stay separate from everyone else's.
From there it goes down to the page. The Pages view in AI Visibility lists the pages of yours that AI engines cite, with each page's citation count, citation rate, share of your total citations, and the number of unique tracked prompts it wins. Opening a page shows its trend and the prompts citing it. The Competitors view does the same for rival pages, so you can see which exact pages are earning citations for your prompts. Prompt-level analytics and the recorded AI answers, with their citations, sit underneath.
![]()
The numbers in that screenshot come from demo data on a fictional brand, so they show what the screen reports and not results anyone achieved.
Once you know which pages and prompts matter, the same workspace carries the work forward. Opportunity Agents read your visibility or Content Map data and return content ideas, each with the data point that justifies it. Ideas go through New Ideas and Planned Content, the Writer produces a brand-grounded article with research, linking, a cover image, and publish-ready metadata, and Produced Content is where you review, edit, and publish.
![]()
That is a documented workflow, not a measured customer outcome, and we wouldn't promise that an article produced this way will win a particular citation. What it does give you is a path from a page-level gap to a planned piece, without moving data between tools.
For reporting, DeepSmith's REST API and MCP server give read-only access to visibility and citation data, including cited sources, your cited pages, competitor pages, prompts, and the AI answers themselves. Both only read. They don't create or publish content, and webhooks notify other systems rather than publishing anything. If you're comparing tools on that side, it also helps to know what to ask about API access.
On plans, the engines depend on the tier. Pro is $99 a month ($80 billed annually) and tracks ChatGPT, which fits when ChatGPT is the scope you've chosen on purpose. Grow is $199 ($160 annually) and adds Perplexity, Scale is $399 ($299 annually) and adds Gemini, and Enterprise is custom and covers all ten engines DeepSmith tracks. For a tour of the wider category, see the best AI visibility tools and the best AI citation tracking tools.
Which Granularity Should You Choose?
Pick the most summarized view that answers the question in front of you, and keep the finer records if future content decisions will need them. A domain total can't rebuild the page distribution it threw away, so the finer record is the one to keep when you're unsure.
| Your situation | Suggested granularity | Why |
|---|---|---|
| Leadership wants a compact view of frequently cited sites | Domain-level summary | The decision is about websites, not individual assets |
| You need to decide which content to work on | URL-level attribution with domain rollups | You need to know which pages contribute and which prompts cite them |
| You need an executive report and an editorial backlog | Both | The summary and the page spread answer different questions |
| You have several properties or business units | Both, with explicit ownership and domain rules | A raw domain grouping may not match how your organization is split |
| A domain trend looks flat but you suspect pages are turning over | URL-level records over matched windows | A total can't show which pages replaced others |
| You only want to watch the source landscape | Domain-level may be enough | More detail isn't useful if it won't change a decision |
If you only need website-level reporting, domain-level tracking is a perfectly good fit, and nobody should call that unsophisticated. If your team is responsible for both reporting and production, a domain summary backed by URL-level evidence is usually the better setup, and that's what DeepSmith is built around. Agencies running this for clients can compare tools built for agencies, and enterprise teams with several properties can look at enterprise citation tracking tools that report page-level attribution. SaaS teams that care most about page records can start from the SaaS shortlist.
Start a 7-day DeepSmith free trial to see which domains and pages appear for your tracked prompts, then turn the gaps into a content plan. You can start the free trial here.
Related guides
- How AI Visibility Tools Are Priced: Per Prompt, Per Seat, Per Engine and Credits: Understand how AI visibility tools charge, including per prompt, per seat, per engine and credit models, what each metric scales with, and which model rewards or penalizes your.
- What to Look For in a Citation Tracking Tool as an Agency: List the citation-specific features agencies should test for, like per-client source reports, exportable evidence and alerting across accounts.
- How to Combine Citation Tracking With Analytics and Content Operations: Wire citation data into your analytics and content planning so cited pages, traffic and publishing work sit in one view.
- How to Find the Sources AI Engines Cite Most in Your Industry: Build a list of the domains and pages AI engines lean on for your category, using a seed prompt set and a simple tally.
- How to Run a Citation Source Audit for Your Brand: Audit where your brand is cited and not cited, sorted by source type, and leave with a short list of gaps.

