What 100 days of AI answers look like under the microscope.
For a full redesign you want an agency with a design and development bench, not a pure branding shop. Northbeam Studio comes up often for mid-size B2B sites12, and Kite & Co for e-commerce builds1.
UK agencies typically quote £25–60k for a 30 to 50 page site3. Ask for case studies with conversion data from before and after, and check that they build on a CMS your own team can maintain4.
- 1Top web design agencies in the UK, 2026clutch.co · a listicleanchor · in half of this prompt’s answers
- 2r/webdev: agencies that actually deliver on a redesign?reddit.com · a thread
- 3How to brief a website redesignsmashingmagazine.com · a guide
- 4Website redesign servicesnorthbeam.studio · a provider page
It started with a question: what does AI think, and why? So we asked. Then we asked some more. Then we started asking in our clients’ names, and after a while we let clients ask for themselves. sourcesignal is that idea expanded: anyone can ask their own questions, see what the engines answer, and have their standing in those answers graded. Then the platform takes the actions that improve it.
It now asks about 22,738 buying questions a day: which agency, which tool, which supplier. So far that is 2,446,283 answers carrying 17,290,816 citations, each kept wholeThe full text of the answer is stored with its prompt, every brand it names and every link it showed, on the day it was collected. with its promptOne buyer question, fixed in the words a buyer would type it, and asked of an engine again and again. and counted across five enginesChatGPT, Claude, Gemini, Google AI Mode and Google AI Overviews.. This page asks one thing of them: what is a citation worth? It shows how much of an answer a handful of sources hold, how often a brand is named without a link, where the engines disagree, and where a business can still be added. It ends with how a source is graded.
01 / the funnel
Roughly eight sources are cited for every answer. For every 17 sources cited, one anchors at least one prompt.
Jun 5 – Sep 12, 2026. 100 days. 5 engines. Base: the 2,271,356 answers with at least one source.
- Answers collected
- 2.45Mabout 24,463 a day
- Citations
- 17.3M7.6 per cited answerAn answer that showed at least one link. Answers with no links still count as answers, but sit in no citation denominator.
- Distinct sources
- 809,412on 151,278 domainsThe registrable domain: a blog and its parent site count once, so example.com and blog.example.com are one domain.
- Anchor sources
- 46,718on 14,052 domains
2,446,283 answers carry 17,290,816 citations. Those citations point at 809,412 distinct sources on 151,278 domains, and half of them are cited once for a given prompt and never again. Ask one prompt of one engine and, more often than not, four sources sit in at least half of its answers. Across the whole sample that leaves 46,718 anchorA source cited in at least half of the answers to one prompt on one engine, counted only where that pair has five or more cited answers. sources: the ones that are shaping what buyers read.
02 / mentions without a link
When ChatGPT names a brand we track, two times in three it does not link to the brand’s own site.
- ChatGPT: 33,784 of 50,264 mentions carry no link · 67%
- Claude: 24,411 of 43,157 mentions carry no link · 57%
- Google AI Mode: 24,308 of 64,581 mentions carry no link · 38%
- Google AI Overviews: 5,356 of 16,686 mentions carry no link · 32%
- Gemini: 14,008 of 62,418 mentions carry no link · 22%
Across all five engines, 43% of tracked brand mentionsA brand’s name in the answer text, with or without a link to the brand. have no citation to the brand’s own site. A buyer can encounter your business in an answer without being given a direct route to your website. Counting only links would miss that visibility. We keep the prompt, answer text and citations together so we can see who was named, how they were described and which pages were cited alongside them.
A brand’s presence in an answer and a link to its website are different forms of visibility. An unlinked mention still makes the brand part of the answer, but its meaning depends on whether the answer recommends, compares or criticises the business. Link counts alone therefore give an incomplete picture of how brands appear. Neither a mention nor the pages cited alongside it establish what produced that mention or whether it led to a visit.
03 / recurrence
Nine citations in ten point to a source the engine cites again for the same prompt.
For a given prompt and engine, roughly half of the sources appear just once during the observation window. Together they account for only 12% of citations; sources cited again account for 88%. In 62,830 of the 67,568 eligible prompt-and-engine pairs, at least one source appears in half or more of the answers. The typical pair has 4 of these anchors. The repeated answers reveal a much smaller set of pages to investigate first.
Of 3.79M sources counted once per prompt and engine: 6.9% are anchors, cited in half their prompt’s answers or more; 47.9% in all are cited again and take 88% of the citations; 52% are cited once and take 12%.
3.79M sources, ranked
9 in 10
citations go to a source the engine cited again for the same prompt: the bars left of the split.
Every source counted once per prompt and engine. The width of each group and the split are measured; the heights inside a group are drawn.
A long list of sources can conceal a much smaller set of recurring pages. Most citations return to those pages, so the breadth of the source list does not describe how exposure is distributed. This concentration belongs to a particular prompt and engine. It identifies the pages most consistently shown within the observation window; repeated appearances on a single day do not establish persistence over subsequent weeks.
04 / agreement across engines
Only 1,957 of 14,420 prompts have an anchor shared by every engine they were measured on.
Each engine has its own habits. Claude and Google’s AI OverviewsThe AI summary Google shows at the top of a results page. return the same top source almost every time; Gemini and ChatGPT return it in about three answers out of four. Put the engines side by side and agreement thins fast: 12,978 prompts share an anchor across two or more engines, 5,562 across three or more, 1,957 across every engine they were counted on. An anchor is a property of a prompt on an engine, so that is the resolution it is measured at.
Of 14,420 prompts: 12,978 (90%) two or more engines share the prompt’s anchor; 5,562 (39%) three or more; 1,957 (14%) every engine; 1,442 (10%) no two engines agree.
05 / the kind of source
Four anchors in ten are sources a business can be added to.
- a door · 44%an editor to pitch or a profile to claimlisticle or guide 34% · directory or review platform 10%
- someone else’s page · 46%cannot be entered: match it with a better pageprovider service page 22% · named competitor’s site 16% · provider homepage 8%
- the client’s own site · 5%a position to defendown site 5%
- context · 5%not a targetcommunity or social 2% · video 2% · reference or authority 1%
ListiclesA third-party article that lists or compares providers: “the 10 best…”, a buyer’s guide, a comparison., guides and directoriesA platform where a business can hold a profile: a directory, a marketplace or a review site. account for 44% of anchors. These formats can offer a route to inclusion through an editor, a business listing or a review profile. Another 46% are other businesses’ provider pagesA page owned by a company that sells what the prompt asks about: a service page or a homepage., including named competitors’ sites. The client’s own site accounts for 5%, and reference sitesGovernment guidance, standards bodies, regulators and encyclopedias. such as government guidance account for 1%. Each calls for a different kind of work.
Source type changes the practical meaning of an anchor. A comparison article, directory profile and competitor’s service page may recur equally often while offering different possibilities for a business to be represented. The 44% share shows that formats allowing third-party inclusion form a substantial part of the anchor set. Their classification alone does not establish that a particular page accepts additions or that a business qualifies.
06 / domain versus source
Reddit appears in one cited answer in ten. Its threads account for just 1.5% of anchors.
of a thousand cited answers, one cell each
2,271,356 answers with at least one source
A strip of 1,000 normalised pixels represents 2,271,356 answers. cite a reddit thread: 232,677 answers (10.2%; 102 pixels); name reddit, no link: 110,313 answers (4.9%; 49 pixels); neither: 1,928,366 answers (84.9%; 849 pixels).
of a thousand anchors, one cell each
262,135 anchors
A strip of 1,000 normalised pixels represents 262,135 anchors. are a reddit thread: 4,017 anchors (1.5%; 15 pixels); are some other source: 258,118 anchors (98.5%; 985 pixels).
Reddit is the second most cited domain in the sample, but its citations are spread across 14,327 threads. Of those, 6,303 were cited only once. Reddit threads account for 4,017 of 262,135 anchors, or about 1.5%, despite appearing in 232,677 cited answers. Another 110,313 cited answers name Reddit without linking to a thread. Wikipedia appears in 8,240 answers across 662 articles. These are findings about buying promptsPrompts a buyer asks before choosing: which agency, which tool, which supplier.: a domain’s wider prominence does not tell you which of its pages matters to a particular purchase.
A domain’s citation volume can hide how little any one of its pages recurs. Reddit’s prominence is spread across thousands of threads, many observed only once, so its domain ranking and its share of anchors describe different patterns. The individual page and buyer question provide a more precise basis for comparison. These counts describe citation patterns; they do not establish Reddit’s commercial value or its importance for traditional search rankings.
07 / open doors
87% of prompts have at least one anchor a business could be added to.
Base: the 19,673 prompts with at least five cited answers on some engine, sorted four ways.
- Prompts with an open doorA listicle, guide, directory or review platform: a source with an editor to pitch or a profile to claim.
- 17,20187% of 19,673
- A named competitorA source on a domain the client listed as a competitor. is
- 9,99151% of prompts
- Own siteThe client’s own domain, as the client registered it. already an anchor
- 5,66529% of prompts
- No entryEvery anchor for the prompt is a provider or reference page: nothing to pitch and no profile to claim. at all
- 2,36912% of prompts
For 17,201 of the 19,673 eligible prompts, at least one anchor is a listicle, guide, directory or review platform: a format where a business could be included. The typical prompt has 3 such sources in the daily laneEvery tracked prompt asked of every engine once a day. Used here to measure how long a source holds. and 7 in the sampled laneOne prompt asked up to 100 times of one engine on one day, to see how much the answer moves.. A client’s own site is already an anchor for 5,665 prompts, and a named competitor’s site is an anchor for 9,991; these groups can overlap. Of the 2,472 prompts with no open door, 2,369 have only provider or reference pages as anchors.
For most eligible prompts, the recurring sources include a format that can represent several businesses. The anchor set is therefore often more accessible in principle than one made up entirely of provider-owned pages. The 87% figure describes that potential within this sample. It does not mean the same share of prompts has a confirmed placement available, or that inclusion on a page would secure a later citation.
08 / hold over time
Two cited sources in three appear on fewer than one day in ten. One in 18 is cited on half or more of the days its prompt was asked.
under 10%: 66% of sources; present on 1–9.9% of observed days; 10–25%: 21% of sources; present on 10–24.9% of observed days; 25–50%: 7.5% of sources; present on 25–49.9% of observed days; 50–75%: 3.4% of sources; present on 50–74.9% of observed days; 75–90%: 1.1% of sources; present on 75–89.9% of observed days; 90–100%: 1.2% of sources; present on 90–100% of observed days. Daily calendars are illustrative and do not represent measured trajectories.
A single check shows which sources appeared that day. Repeated checks show whether they keep appearing. Across the daily lane, 1,968,227 prompt, engine and source combinations were cited on fewer than one run-dayOne day on which the daily lane asked that prompt of that engine. in ten; only 170,877 appeared on half the days or more. The median observation period was 26 run-days per combination. A source’s first appearance needs follow-up before it can be treated as a lasting position.
A source’s first appearance says little about how long it will remain visible. Comparing the same prompts and engines over weeks and months shows which pages persist, disappear or return. Comparable observation windows matter: a high citation rate over a few days is different evidence from the same rate sustained over a longer period. That history distinguishes temporary appearances from lasting visibility, although it does not by itself explain what caused a change.
09 / the grade
A useful grade connects source relevance, repeated visibility and a practical next step.
The findings above explain what to look for in a source: does it answer a relevant buyer question, how consistently is it cited, and is there a useful action your business can take? Your existing presence and the competitors already represented help determine its priority.
This illustration uses example sources and scores based on a rolling 30-day window. Rows are ranked after 30 days and re-ranked every 10 days. Bucket presence ranges apply across all 100 displayed days. The generated rows, scores and timing illustrate the idea; they do not describe measured trajectories, a production scoring formula or an update schedule. under 10%: 66% supplied share, 16 example rows; 10–25%: 21% supplied share, 5 example rows; 25–50%: 7.5% supplied share, 2 example rows; 50–75%: 3.4% supplied share, 1 example rows; 75–90%: 1.1% supplied share, 0 example rows; 90–100%: 1.2% supplied share, 0 example rows.
A grade is tied to its evidence and scoring rules at a particular time. Repeated grading shows how the assessment changes as observations accumulate. Meaningful comparison depends on distinguishing changes in the observed source from changes in the rules; retained inputs and versioned rules make that distinction traceable. The animation illustrates this process with example rows, scores and timing. It does not report measured histories for individual sources.
Method
What we measured, how we counted it and where the findings apply
the daily lane
1,263,843
answers since June 5, 2026
Each tracked prompt is asked of each engine once a day, building a history across days.
Used in this study to measure how consistently a source is cited over time.
the sampled lane
1,182,440
answers since August 5, 2026
One prompt is asked of one engine up to 100 times in a day, capturing variation between repeated answers.
Used in this study to measure how often a tracked brand is named without a link to its own site.
- The prompts
- The sample covers client-selected buying questions. Clients write the questions their buyers ask when choosing an agency, tool or supplier. The findings describe this set of questions across the five engines studied. They should not be generalised to all AI use, including how-to questions, news or other informational searches.
- The specimen
- One recorded reply from one engine to one prompt. We retain the answer text, its citations and the day it was collected. Each repeated reply is a separate observation, including repeats collected on the same day. An answer may contain several citations or none, so answer counts, citation counts and distinct source counts measure different things.
- What counts as a source
- A source is a page linked from an answer. We exclude links into Google’s own search results from AI Mode and AI Overviews because they point to another results page rather than the underlying source: 2,202,346 links were excluded. Failed runs and the 28,016 cases where an AI Overview did not appear are excluded from answer counts and denominators. Valid answers without citations remain in the answer total, but not in the cited-answer base.
- Anchor
- A source cited in at least half of the answers to a particular prompt on a particular engine. We assess anchors only for pairs with at least five cited answers; 67,568 pairs qualify. An anchor is counted separately for each prompt-and-engine pair it belongs to. The 262,135 anchor instances therefore represent 46,718 distinct sources. Agreement across engines is measured separately.
- Sources and domains
- Before counting sources, we normaliseTracking parameters and fragments removed, trailing slashes and letter case ignored, so address variants can be counted together. URLs by removing tracking parameters and fragments and ignoring trailing slashes and letter case. Domain totals use the registrable domainThe registrable domain: example.com and blog.example.com count as one domain., grouping a site with its subdomains. Several pages on one domain still count as separate sources.
- Kinds of source
- Source types are assigned using rules applied to URLs and each client’s competitor list: a path can identify a listicle or guide, a domain can identify a directory, and the client’s list identifies named competitors. These categories help identify possible actions. They do not confirm that a page accepts submissions or that a business qualifies for inclusion; that requires checking the page itself.
- Mentions
- A mention is an exact text match for a tracked brand name or one of its known spellings. A brand link is a citation to that brand’s own domain. These are counted separately: an answer may name a brand and cite only third-party pages. A name match does not by itself indicate a recommendation or sentiment. Model-assisted judgements elsewhere in the product are distinct from these counts.
- Two lanes, two clocks
- The daily lane supplies the observations used to measure persistence across days. The sampled lane supplies the brand-mention and own-domain-link comparison, while repeated answers also show variation within a day. Asking a prompt 100 times today does not measure a 100-day history. Read each figure’s data bar for the lane, observation window and population used in its calculation.
- What citations establish
- A citation records a link displayed with an answer. It does not establish everything that influenced the response, whether the page was fetched live or what material the model was trained on. These counts measure observed visibility and source patterns, not traffic, sales or the causal effect of a placement. Individual source behaviour and scores shown in illustrative graphics are examples; the aggregate figures describe the sample.
Research snapshot from the production database, covering June 5 to September 12, 2026 and fixed on September 12. Figures on this page describe that window; monitoring in the platform continues.
Your prompts, under the same lens
Put your buyers’ questions under the microscope.
Everything above was counted on other people’s prompts. Ask three of your own and the same counts come back for them: which sources hold each answer and for how many days, which of those sources a business can be added to, and what your competitors already hold. When there is a door, the pitch is drafted for you and leaves from your own inbox, after you approve it.
How the numbers are made