Beacon · agency deck
Content pass. Black and white, no design yet. Structured as situation, complication, resolution,
with an action title on every slide: reading the forty titles below in order should reproduce the whole argument
without opening a single slide. If it does not, the flow is wrong and we fix it here rather than in the artwork.
Governing thought
- Knowing what the assistant says about your brand is the easy third of the job. The other two thirds change it.
Situation · where the customer now asks
- Your customer stopped searching and started asking, and the answer arrives without a list of links.
- In Australia this is already mainstream: more than 70% of 25 to 49 year olds use AI tools.
- The assistant is already inside the purchase, not just the research.
Complication · the answer layer is not yours, and it moves
- Four platform decisions in twelve months rewrote which brands get named, and no brand caused any of them.
- Google now answers with Google: its own domain went from 0.4% to 7.3% of AI Mode citations in five months.
- Reddit was effectively delisted twice, and brands that had built their visibility there lost it overnight.
- One model release cut the number of citation slots per answer by 27%, across every model at once.
- Since August, ChatGPT reads your own website directly, so your site is now a retrieval surface.
- The brands affected by all four events could not see any of it happening.
Complication · and the market's answer is a dashboard
- Twenty-one tools now sell AI visibility, and measurement stopped being a differentiator this year.
- They price the same thing differently, and one of them renames fifty prompts as six thousand responses.
- Every one of them stops at the same place: a chart, and a recommendation somebody else has to act on.
Resolution · three layers, and we run all three
- There are three ways into an AI answer, and the category sells only the one that describes it.
- Beacon runs the layer that changes the answer, the layer that measures it, and the layer that buys the gap.
- Run in order, each layer makes the next one cheaper.
Resolution · the measurement layer, eleven modules
- Beacon reports on eleven modules, tracked daily across nine assistants.
- Share of voice carries over from media into the assistant, per engine and per market.
- Every answer is built from about five sources, and we show you which five.
- Product pages became the most cited format this year, on two engines independently.
- We keep what the model actually said, so a wrong claim can be produced rather than described.
- The same measures run against a named competitor set, on the same prompts on the same days.
- Crawler logs are the earliest signal there is, and analytics tags cannot see them.
- We start a client with twelve months of history instead of day one.
- Assistant traffic lands as direct in standard analytics, so we measure it on our own script.
- One customer question becomes several machine queries, and those are what your content has to match.
- The same channel behaves differently on every engine, so one social strategy will misallocate.
- Paid placement inside answers went from zero to a third of responses in three months.
- Every finding lands as assigned work, and one column of it only we can generate.
Resolution · the layer that changes the answer
- There are three routes in, and only one of them persists after you stop paying.
- Brands are live in market today, and two of them can be installed in the next five minutes.
- A connector is a server the brand controls, configured once per platform.
- Philips answers as Philips, with confirmed specifications and live pricing.
- The same connector answers the questions that arrive after the sale.
- Four things separate a branded answer from a model guessing.
- One connector carries six different surfaces, and adding one is configuration.
- An approval does not have to be bought again next quarter.
Resolution · buying the gap, and what it costs
- Paid becomes a target list instead of a keyword guess.
- The second brand costs a fraction of the first, and Philips already paid for the path.
- Three phases to stand up, and reporting can be bought on its own.
BeaconEVAA Pte. Ltd.
Beacon
Branded answers on ChatGPT, Gemini and Claude. Measured, and moved.
BeaconGoverning thought
Knowing what the assistant says about your brand is the
easy third of the job. The other two thirds are what change it.
Downstream · measure
What the assistant says today
Visibility, citations, sentiment, competitors, crawler behaviour. Twenty-one tools now sell this. It is table stakes.
Upstream · change
What the assistant says tomorrow
The brand's own data connected to the platform, so the answer comes from the brand rather than from model memory. Built and accepted with Philips.
Paid · close the gap
The questions still going elsewhere
Advertising inside the answer, bought against the exact prompts the reporting shows are lost.
This deck is organised around that sentence. Section 01 is the problem, section 02 is why the current
answer to it is insufficient, and everything after that is the three layers.
BeaconSituation
Section 01
Discovery moved, and it took the answer with it.
The behaviour is already mainstream in Australia. What follows is not a forecast.
BeaconSituation
Your customer stopped searching and started asking, and the answer now arrives
without a list of links.
A search results page offered ten links, ads and a chance to be compared. An assistant returns one
synthesised answer naming a handful of brands. Everything outside that handful is invisible, regardless of how it ranks.
Then
Ten blue links
The customer saw the field and chose from it. Position mattered, and everyone on page one got a hearing.
Now
One answer, five sources
ChatGPT cites around five sources per answer. Google AI Overviews and Perplexity cite about ten. That is the entire inventory.
Consequence
Being ranked is not being named
A brand can hold position three on Google and never appear in the answer the customer actually reads.
This is the shift the rest of the deck is about. It is not a channel being added. It is the layer
where the shortlist gets written moving somewhere the brand has no presence.
Source · Promptwatch public data, captured August 2026
BeaconSituation
In Australia this is already mainstream: more than 70% of 25 to 49 year olds use AI tools.
The highest rate of any age group, and it maps almost exactly onto the buying population for a
considered household purchase.
>70%
of Australians aged 25 to 49 use AI tools, the highest rate of any age group
10.5M
ChatGPT users in Australia, 45% of the AI audience
One platform carries most of the audience, which makes the first connector build a
straightforward decision rather than a portfolio one.
Source · Roy Morgan, AI tools usage, March 2026
BeaconSituation
The assistant is already inside the purchase, not only the research.
Influence at the point of decision is there whether or not the brand has anything to say inside it.
57%
say AI influences their shopping choices
41%
pay attention to AI generated summaries
39%
use AI to help make buying decisions
31%
have already acted on an AI suggestion
Only 29% say they trust AI summaries. The gap between 57% influenced and 29% trusting is the
commercial opening: customers want the answer confirmed by the brand, and a branded answer is the only format that does that.
Source · GWI, InsideRetail, Adyen
BeaconComplication
Section 02
The answer layer belongs to the platforms, and they rewrite it without warning.
Four documented events in twelve months. Not one of them was caused by anything a brand did to its website.
BeaconComplication
Four platform decisions in twelve months rewrote which brands get named, and
no brand caused any of them.
Each of these is dated, documented and measurable. Each changed the composition of AI answers within
a single day.
September 2025
Reddit changes its access terms
Reddit's share of ChatGPT citations fell from 10 to 14 per cent down to roughly 1 per cent within days. It did not recover.
4 March 2026
GPT-5.3 ships
Sources cited per answer dropped from 6.4 to 4.7. A 27 per cent cut in available slots, across every ChatGPT model at once.
8 August 2026
ChatGPT starts reading domains directly
Site-scoped queries went from 1 in 280 to 1 in 6 overnight, and total searches per answer nearly doubled.
14 August 2026
Reddit falls again
From 3.83 per cent of ChatGPT citations to 0.52 per cent in a single day, an 86 per cent drop.
A brand that had built its AI visibility on Reddit lost it twice, for reasons entirely outside its
control, and had no instrumentation that would have told it on either occasion.
Source · Promptwatch public data, captured August 2026
BeaconComplication
Google now answers with Google: its own domain went from 0.4% to 7.3% of AI Mode citations in five months.
Counting YouTube as well, Google-owned properties account for more than one in ten citations in AI
Mode, and roughly one in fifteen in AI Overviews. There is no equivalent behaviour on ChatGPT.
| google.com share | AI Overviews | AI Mode |
|---|
| January 2026 | 0.37% | — |
| March | 0.89% | 0.58% |
| April | 0.61% | 0.71% |
| May | 1.15% | 4.24% |
| June | 2.30% | 7.31% |
What it means commercially
Your Google Business Profile is now a visibility asset
When AI Mode answers with Google's own surfaces, the way to be in that answer is to be inside those surfaces.
Business Profile, Maps and Merchant Center stop being local search housekeeping and become part of the same
programme as the website. Most brands still have them filed under a different budget and a different agency.
By June, google.com alone out-cites YouTube and Reddit combined in AI Mode.
Source · Promptwatch public data, captured August 2026
BeaconComplication
Reddit was effectively delisted twice, and the brands built on it
lost their visibility overnight.
Worth dwelling on, because Reddit was the single largest citation source on ChatGPT on both occasions.
This is what platform dependency costs.
September 2025
10–14% down to about 1%
Reddit changed its access terms. A roughly 90 per cent fall in days. What had been the dominant citation
source on ChatGPT stopped being one.
May to June 2026
6.11% back down to 3.71%
Recovered to become the number one cited domain again, peaking above every other source combined, then lost
40 per cent of that share in a month.
14 August 2026
3.83% down to 0.52%
An 86 per cent fall in one day. Google's engines showed nothing comparable, which rules out anything the
brands or Reddit did to the content itself.
The lesson an agency can sell: visibility rented from a third-party platform can be withdrawn by a
decision neither you nor your client is party to. Visibility on data the brand controls cannot.
Source · Promptwatch public data, captured August 2026
BeaconComplication
One model release cut the number of citation slots per answer by 27%, across every model at once.
GPT-5.3 shipped on 4 March 2026. Within a day, the number of sources ChatGPT cited per search-enabled
answer fell and stayed down. A month later there was no recovery.
6.4
average sources cited per answer in the week before the rollout
4.7
average by late March, and still there a month later
27%
fewer places for any brand to appear, overnight
1 day
how long the transition took, across GPT-5.3, 5.4 and 5-Mini simultaneously
The brands squeezed out first are the marginal ones: cited sometimes, never first choice. That is
most brands. And any of them looking at a traffic dip in March would have blamed their own content.
Source · Promptwatch public data, captured August 2026
BeaconComplication
Since August, ChatGPT reads your own website directly, which makes your site
a retrieval surface.
On 8 August 2026, domain-scoped searches went from roughly 1 in 280 of ChatGPT's background queries to
1 in 6, in a single day. Total searches per answer nearly doubled at the same time, which proves these are additional
searches rather than replacements.
Before
Search the open web, see what comes back
The assistant looked for pages about the category and the brand appeared, or did not, depending on what
happened to rank.
After
Go to the brand's own site and look
The assistant now runs the equivalent of a site-scoped search against your domain for the topic it is
answering. It reads what is there.
What it costs you
Structure is now revenue
Thin category pages, unindexed sections and broken internal search stop being an SEO nuisance and start
costing answers. And a site-scoped search only surfaces what is indexed.
This is the change that connects technical site quality directly to whether the assistant can
describe the brand accurately. It is also the change that makes a connector the cleaner solution, because a
connector answers from structured data rather than from whatever the crawler could parse.
Source · Promptwatch public data, captured August 2026
BeaconComplication
The brands affected by all four events could not see any of it happening.
Each event is visible in the aggregate data after the fact. None of it appears in the reporting a
brand or its agency actually receives.
Why analytics missed it
The traffic arrives unlabelled
Assistants strip the referrer, so sessions from an AI answer land in analytics as direct traffic. A brand
losing assistant visibility sees an unexplained dip in direct, months later, with nothing to attribute it to.
Why the crawler data missed it
Bots do not run JavaScript
Standard analytics is a script that executes in a browser. AI crawlers do not execute it. The visit that
determines whether you can be cited at all is invisible to the tool most brands rely on.
So the sequence for a brand in 2026 has been: lose visibility to a platform decision, see it as an
unexplained dip in direct traffic one or two quarters later, and have no way to attribute it. That is the problem
Beacon's measurement layer exists to close, and it is only the first third.
BeaconComplication
Section 03
The market's answer to this is a dashboard.
Twenty-one tools now sell AI visibility. They measure the problem accurately and none of them change it.
BeaconComplication
Measurement stopped being a differentiator this year.
These are the category's own numbers, published by one of the vendors inside it as part of its
competitive comparison.
21
tools now selling AI visibility tracking
18
track prompts and citations
12
read server or CDN crawler logs
10
detect the brand named inside third-party pages
Their own comparison page opens with the line: “Everyone measures. Almost nobody ships the
fix.” That is a vendor describing its own category, and it is the most useful sentence written about this
market so far.
Source · Promptwatch public data, captured August 2026
BeaconComplication
They price the same thing differently, and one of them renames
fifty prompts as six thousand responses.
A prompt is one question tracked. A site is one brand tracked. Everything else on these pricing pages
is presentation.
| Promptwatch | mentions.so | Profound |
|---|
| Entry price | $95 per month | $49 per month | $99 per month |
| What that buys | 50 prompts, all nine engines, API and MCP included |
25 prompts, one site, three engines, unlimited seats on every plan |
50 prompts on ChatGPT only. Three engines costs $399 |
| Top tier | $579 | $399 agency, unlimited sites, white label | Enterprise, quoted |
| What happens next | An agent writes a page into Webflow or Framer |
You get a prioritised list | Publishes into WordPress, Drupal, AEM |
Two things an agency should take from this table. Per-seat and per-domain pricing is the trap, and
the cheapest per-response vendor is already at one cent, so there is no margin left in reselling measurement alone.
Source · Vendor pricing pages, August 2026
BeaconComplication
Every one of them stops at the same place: a chart, and a recommendation
somebody else has to act on.
What the category delivers
Evidence, then a handoff
A number, a trend line and a recommendation. Someone client-side then has to interpret it, brief it, write
it, publish it, and wait a retrieval cycle to find out whether the answer changed. In practice that work
competes with everything else in the marketing calendar and usually loses.
What actually moves an answer
Change what the model can retrieve
Either the model finds something different about the brand, or the brand is connected to the platform so the
answer is generated from its own data. Publishing another article is the slowest and least reliable of the
available routes, and it is the only one the category offers.
One vendor in this category routes readers to an outside GEO agency inside its own blog post, for
the strategy and resourcing work the tool does not cover. That is the gap, admitted in writing, on their own site.
BeaconResolution
Section 04
There are three layers. The category sells one. We run all three.
Measurement is the layer that describes the answer. It is necessary, it is table stakes, and on its own it changes nothing.
BeaconResolution
There are three ways into an AI answer, and the category sells only
the one that describes it.
They are not alternatives to each other. They have different timelines, different ceilings, and they work in sequence.
Upstream · changes the answer
Connectors and apps
The brand's own data linked to the platform, so the assistant answers as the brand with confirmed facts,
live pricing and native actions. Live on ChatGPT, Gemini and Claude in every market. Persists until the
customer turns it off. Built and accepted with Philips.
Downstream · measures it
Reporting
What the assistants say now, which sources they pull from, how that compares against named competitors, and
which pages the crawlers reached. Eleven modules. This is the layer the twenty-one tools sell, and we match all of it.
Paid · buys the gap
Advertising inside the answer
Ads now appear in roughly a third of ChatGPT responses that ran a search. We buy against the exact prompts
the reporting shows are being lost, with a known competitor on the other side.
Read left to right, that is a service. Read only the middle column, that is a subscription.
Source · Promptwatch public data, captured August 2026
BeaconResolution
We match the category on every platform and every integration, and we
own the two layers they do not have.
Feature parity is deliberate. Nothing on a competitor's checklist should be a reason to choose them,
which leaves the argument where it belongs.
| Monitoring tools | Enterprise platforms | Beacon |
|---|
| Engines tracked | 3 to 9, often as paid add-ons | Tiered by contract |
All ten, on every plan. ChatGPT, AI Overviews, AI Mode, Gemini, Claude, Perplexity, Copilot, Grok, DeepSeek, Llama |
| API and MCP | Gated above entry, or absent | Enterprise only |
Both, on every paid plan. The data goes wherever the agency already works |
| Action queue | A prioritised list | Some, with owners |
Kanban board, cards assigned, moved through to done, with a connector column |
| Crawler logs | 12 of 21 read them | Yes |
Yes, plus historical backfill at onboarding |
| White label | Top tiers only | Enterprise |
Included, with a client portal and scheduled reporting |
| Changes the answer | No | Writes a page and publishes it |
Connector. The answer changes when the brand's data changes |
The last row is the whole argument. Everything above it is the price of being taken seriously.
BeaconResolution
Run in order, each layer makes the next one cheaper.
This is the sequence we ran on Philips, and it is why the three layers are one service rather than three products.
First · measure
Find out which questions matter
Reporting shows which questions the category is asked, which brands the assistants name in reply, and which
of those answers are wrong, stale or missing the brand entirely.
Second · fix at the source
Turn the losses into capabilities
The questions worth owning become connector capabilities. Specifications, current range and live pricing
stop being model memory and start being brand data that updates itself.
Third · buy what is left
Media against a known target
Whatever still goes to a competitor is a media buy with a named prompt, a named rival and a measurable
outcome, rather than a keyword guess.
Each pass narrows the next. The reporting is what makes the media efficient. The connector is what
makes the reporting improve rather than just record.
BeaconReporting
Section 05
The measurement layer, in full.
Ten assistants, tracked daily, at country and city level. Eleven modules. One slide each.
BeaconReporting
Beacon reports on eleven modules, tracked daily across ten assistants.
01
Visibility and share of voice
05
Competitor benchmarking
07
AI traffic and conversion
Delivery
API, MCP, white label
Everything after this slide is one of these in detail. The twelfth tile is how it reaches the
agency: a REST API and an MCP server on every paid plan, plus a white-label client portal.
BeaconReporting · 01
Share of voice carries over from media into the assistant, per engine and per market.
The headline number an agency planner already knows how to brief against, measured somewhere new.
- Ten engines: ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Claude, Perplexity, Copilot, Grok, DeepSeek and Llama.
- Per market and per language, run from the country the customer is actually in, down to city level where it matters.
- Per prompt, so the fifty questions worth owning are visible individually rather than rolled into an index.
- Tracked daily, because platform behaviour changes overnight and a monthly snapshot cannot tell you which day it moved.
Measured on the real interfaces, because an answer from a developer API is not the answer a
signed-in customer sees, and a location passed to an API is treated as a hint rather than a place.
BeaconReporting · 02
Every answer is built from about five sources, and we show you which five.
Citation inventory is the scarcest thing in AI search. Knowing who holds the slots on your prompts is
the difference between a strategy and a guess.
- ChatGPT cites roughly five sources per answer. Google AI Overviews and Perplexity cite about ten. That is the entire inventory per question.
- The most cited single domain on ChatGPT has never held more than about six per cent of citations, and usually holds three to four.
- Around ninety per cent of citations go to the long tail, which is why a specialist brand can realistically win a specific question.
- When a source drops out of an answer we can see which source replaced it, which turns a loss into a brief.
This is also the module that separates "we lost visibility" from "the engine started citing fewer
sources". Same chart, completely different problem, completely different response.
Source · Promptwatch public data, captured August 2026
BeaconReporting · 03
Product pages became the most cited format this year, on two engines independently.
Two separately operated engines moved the same way at the same time, which makes this a property of
how AI search works rather than a quirk of one platform.
33%
of ChatGPT citations now go to product pages, up from 18% in March
16.4%
of AI Overview citations, up from around 9% in January
28 Jul
the day product pages overtook listicles on AI Overviews for the first time
0.05%
of citations go to markdown files, against 99.94% to ordinary HTML
Two things to say to a client out loud. Specifications, pricing and availability on the brand's own
pages are now citable primary sources, which makes this a product-data job before it is a content job. And the
widespread advice to publish markdown copies of your pages for the models is worth nothing: the ratio is two
thousand to one against.
Source · Promptwatch public data, captured August 2026
BeaconReporting · 04
We keep what the model actually said, so a wrong claim can be
produced rather than described.
Sentiment as a score is close to useless on its own. Sentiment attached to the sentence that caused it
is an action.
- Favourability scored per answer per engine and tracked over time, with the score always linked back to its source text.
- The full response kept verbatim, so a discontinued product, a superseded price or a competitor's claim can be evidenced on the spot.
- Perception themes surfaced across answers, for example a pricing objection the model keeps raising unprompted.
- Alerts when an assistant starts saying something materially new about the brand.
Most brands find out a model is telling customers something out of date when a customer repeats it
back to them in a complaint. This module shortens that loop from months to a day.
BeaconReporting · 05
The same measures run against a named competitor set, on the same prompts on the same days.
The assistant names a small number of brands per answer. Benchmarking is the only way to know whether
the client is in that set, and who is there instead.
- Share of voice against the competitive set the client already reports on, so it lands in an existing review rather than needing a new one.
- Which competitor the assistant recommends when the client is not named, and what reason it gives for it.
- Which sources are carrying that competitor into the answer, which is usually the fastest brief in the whole report.
- Movement flagged when a rival gains share on a tracked prompt, rather than discovered at the quarterly review.
For an agency this is the slide that renews the retainer, because it is the one the client forwards
to their own leadership.
BeaconReporting · 06
Crawler logs are the earliest signal there is, and analytics tags cannot see them.
Before an assistant can quote a page, its crawler has to fetch it. Server and CDN logs record whether
that happened and what the bot received.
- Which AI crawlers reached the site, which pages they asked for, and what status code they got back.
- Requests validated against each operator's published IP ranges, so spoofed user agents are filtered rather than counted.
- Meta's indexing crawler went from 2.2% of AI crawler traffic to 37.8% between mid July and 9 August 2026, consistent with Meta building its own search index.
- Anthropic's citation crawler grew more than a hundredfold between December 2025 and April 2026.
Most AI crawlers do not execute JavaScript, and standard analytics is JavaScript. The one visit that
determines whether a brand can be cited at all is invisible to the tool most brands rely on. A robots.txt rule
written in 2023 is quietly excluding some of them right now.
Source · Promptwatch public data, captured August 2026
BeaconReporting · 06
We start a client with twelve months of history instead of day one.
The strongest claim the incumbents make is that log history cannot be bought. That is true of their own
research dataset. It is not true of a client's dashboard.
How the category onboards
The clock starts at signature
A connector is switched on and begins collecting that day. Three months pass before a pattern is readable.
The first quarterly review has nothing to compare against, and that is also the review where the client decides
whether to keep paying.
How Beacon onboards
The logs already exist
Cloudflare, Fastly and Akamai retain them, and they belong to the client. We ingest that history at
onboarding, so the first review opens with twelve months of crawler behaviour already plotted, and the four
platform events from section 02 visible in the client's own data.
This is an ingestion design decision rather than a pitch line, and it has to be built in from the
start. It is also the hardest objection in this category to answer, so it is worth answering first.
BeaconReporting · 07
Assistant traffic lands as direct in standard analytics, so we measure it
on our own script.
This is why AI referral numbers in most client decks are wrong, and wrong in the same direction.
- Sessions attributed to each assistant by name, rather than collapsed into direct traffic along with everything else.
- Conversion and revenue tied back to the prompt that produced the visit, where the platform passes enough to do it.
- Cost per enabled account for connector campaigns, which is the number a media plan can actually be built on.
- Reported monthly against the plan the account is already running, in the format it already uses.
Assistants strip the referrer, so anything read out of a standard analytics install inherits the
undercount. Any vendor reporting AI traffic from a client's existing analytics is reporting the gap rather than the traffic.
BeaconReporting · 08
One customer question becomes several machine queries, and
those are what your content has to match.
The customer types one thing. The assistant searches for several other things. Content briefed against
the customer's phrasing can miss every one of them.
- The exact queries each assistant generates from a tracked prompt, captured per run rather than inferred.
- Query length has roughly halved in a year, from about 117 characters to about 53. Assistants now search like a person typing keywords rather than pasting a sentence.
- Since 8 August, roughly one in six ChatGPT queries is scoped to a single domain, so headings and titles have to read like search queries.
- Volume and difficulty scored per prompt, so a prompt set can be prioritised rather than tracked flat.
Practically: a heading like "Best air fryer for a family of five, 2026" matches a fifty-character
machine query. "What should you look for when choosing an air fryer?" does not.
Source · Promptwatch public data, captured August 2026
BeaconReporting · 09
The same channel behaves differently on every engine, so
one social strategy will misallocate.
Tracked as channels in their own right rather than as rows inside a citation table, because the
planning decision is different for each one.
Reddit
ChatGPT's largest source
Around 5% of its citations, more than twenty times its next social source. And 71% of the cited posts had
fewer than ten upvotes.
YouTube
Google's largest source
Above 4% of AI Overview citations. On ChatGPT it is 0.05%, effectively invisible. 80% of cited videos come
from channels under 100,000 subscribers.
LinkedIn
Three different products
ChatGPT cites company pages. Google cites Pulse articles at over 42%. Perplexity cites ordinary feed posts.
Same platform, three optimisations.
X
Not a channel
Never exceeds 0.25% of citations on any engine, including Grok, which shares an owner. Worth saying before a
client funds it.
The finding underneath all four: engagement does not earn citations. Retrieval selects for topical
precision, which is why a five-upvote thread that answers the question exactly beats a front-page post that grazes
it. That reframes the budget conversation away from reach.
Source · Promptwatch public data, captured August 2026
BeaconReporting · 10
Paid placement inside answers went from zero to a third of responses in three months.
ChatGPT served no advertising at all before 27 May 2026. This is a channel forming in real time, and
the reporting tracks who is buying against the client.
32.4%
of search-enabled ChatGPT responses carried an ad, seven day average
43.9%
highest daily rate recorded so far
73%
of those ads sat on organic prompts rather than brand or comparison prompts
8.6%
sat on competitor comparison prompts
Two different products share this name across the category. Most tools connect the client's own ad
account, which tells you what you already spent. We track who else is bidding against the prompts the client cares
about, which tells you what to do next.
Source · Promptwatch public data, captured August 2026
BeaconReporting · 11
Every finding lands as assigned work, and one column of it
only we can generate.
A Kanban board rather than a list. Cards carry a type, a priority and an owner, and move through to
done. Nothing in Beacon terminates in a chart.
Source
A page worth being on
A third-party source the assistants keep citing on your prompts, where the brand is currently absent.
Becomes an outreach or PR brief.
Narrative
A perception to correct
A theme the models repeat about the brand, attached to the verbatim answers that evidence it. Becomes a
content or comms brief.
Technical
A retrieval blocker
A crawler blocked in robots.txt, a page returning an error to a bot, a section the crawlers never reached.
Becomes a ticket.
Connector
A capability to add
A question the brand loses that its own data could answer directly. Becomes a connector capability, and the
answer changes without anyone writing a word.
Backlog, to do, doing, done, with an owner on every card. The fourth column is the one no other tool
in this category can produce, because producing it requires owning the upstream layer.
BeaconConnectors
Section 06
The layer that changes the answer.
Built and accepted with Philips for the Versuni range. Everything in this section is from that build.
BeaconConnectors
There are three routes in, and only one of them persists after you stop paying.
| GEO | Advertising | Branded answers |
|---|
| What it is |
Publishing crawlable content on well-ranked sites to influence what the model says |
Placements inside the assistant, sold by the platform |
The brand's own data linked to the model, so answers are accurate, enriched and specific |
| Closest to | SEO | Paid search | Owned media, inside someone else's app |
| APAC availability |
Affects regional answers, driven by language and the ranking of the publishing sites |
ChatGPT advertising is live in Australia |
Live in every market on ChatGPT, Gemini and Claude. Apps and plugins are country enabled |
| Time to effect |
Three to six months of seeding, then continuous renewal |
Immediate on spend, and it stops when spend stops |
Three to four weeks of platform review, then it persists until the customer turns it off |
All three belong in a plan. Only the third one accumulates.
BeaconConnectors
Brands are live in market today, and two of them can be installed
in the next five minutes.
This is not a roadmap slide. Search the plugin list, install, mention the brand, and the branded answer appears.
Coffee
Starbucks
Live in the US. Launched as an app in April 2026, then moved into the plugin directory.
Airline
Turkish Airlines
Live in APAC, with booking, pricing and rewards miles management. Installable from Singapore today.
Quick service
Burger King, Popeyes, Tim Hortons
Live in the US and Canada, with menu, ordering and deals.
Healthcare
Private MD Labs
Live in APAC, with guidance on supplements and lab results.
Worth doing live in the room. The demonstration answers the "is this real yet" question faster than
any slide can, and it arrives in the client's own account rather than in a screenshot.
BeaconConnectors
A connector is a server the brand controls, configured once per platform.
Not an integration project. A maintained source of truth, plus an experience designed against each
platform's own SDK.
Input
Scattered brand data
Product, pricing, stock, reviews, offers, store locations, support content. Usually sitting in four systems that do not talk to each other.
Process
Structured and maintained
Cleaned, structured and updated live, rather than exported once and left to age until someone notices.
The connector
An always-on secured link
Between that source of truth and the platforms. One per brand, serving every platform it is configured for.
Output
Platform experience
Built against each platform's SDK so the answer renders with images, comparisons and actions rather than as plain text.
Each platform runs its own technical review before a connector goes live. Currently three to four
weeks. We have been through it, which is the part that is hard to schedule the first time.
BeaconConnectors
Philips answers as Philips, with confirmed specifications and live pricing.
Rich experience one, the product catalogue. Built and accepted. Screens from that build drop in here.
Image · catalogue-single.png
Individual listing inside the chat
A single product with specification, price and a link, generated from the connector rather than recalled
from model memory.
Image · catalogue-compare.png
Comparative listing across stockists
Live availability and pricing across retailers, answered inside the conversation rather than sending the
customer back out to a search page.
Screens exist in the Philips build and drop straight in. Held as placeholders for this content pass.
Built and accepted · Philips prototype
BeaconConnectors
The same connector answers the questions that arrive after the sale.
Rich experience two, the guided how-to. This is the half of the customer relationship most brands
cannot currently see, let alone answer.
Image · guide-list.png
Guides offered for a registered product
The assistant knows which model the customer owns and offers the guidance that applies to it.
Image · guide-detail.png
Step by step instructions
Brand-authored instructions rendered in the conversation, with every follow-up still answered as the brand.
Commercially this is the strongest argument for a connector over any content approach. It converts a
support cost into a retained relationship, and it runs on data the brand already has sitting in its manuals.
Built and accepted · Philips prototype
BeaconConnectors
Four things separate a branded answer from a model guessing.
Image · anatomy.png
Annotated branded answer from the Philips build
01 · Attribution
It answers as the brand
The platform labels it, so the customer can see the answer came from the brand rather than from the model.
02 · Accuracy
The facts are confirmed
Capacity, wattage, stock and price pulled live from the brand's database rather than recalled from training data.
03 · Action
The actions are native
One tap into the brand's app, a retailer or product registration. The brand decides which links appear.
04 · Persistence
It stays in the conversation
Follow-up questions keep answering as the brand, with no re-asking and no re-finding.
Built and accepted · Philips prototype
BeaconConnectors
One connector carries six different surfaces, and adding one is configuration.
The build cost sits in the first surface. Everything after it is content work against a link that
already exists.
Products
Range and specification
Current line-up, confirmed specifications, comparisons, and what replaced a discontinued model.
Pricing
Live and local
Current price, promotions and where it is in stock, answered per market.
Locations
Where to buy
Stockists and service centres, answered against the customer's actual location.
Support
After the sale
Manuals, troubleshooting, parts, warranty and product registration.
Content
Brand expertise
Recipes, guides, technique. The material that earns the follow-up question and the return visit.
Offers
Promotions and loyalty
Campaign mechanics and membership, surfaced when they are relevant to the question being asked.
For a brand with several categories this is the compounding argument. The second category is a
content exercise against infrastructure that is already standing.
BeaconConnectors
An approval does not have to be bought again next quarter.
This is the argument that separates a connector from a campaign, and it is the one that survives a
budget review.
Media
Rented attention
Impressions stop when the spend stops. Presence evaporates, the audience has to be bought again, and the
next quarter starts from zero with the same brief.
An enabled connector
Retained presence
Every approval persists until the customer turns it off. The brand sits inside the assistant they already
use daily, at close to zero marginal cost, and the platform remembers the relationship between sessions.
$2.11
target cost per enabled account, against a benchmark of roughly $3.50
Modelled on a 95,000 approval launch target. The number that matters to a planner is not the cost
per install, it is that the denominator keeps working after the campaign stops.
BeaconCommercials
Paid becomes a target list instead of a keyword guess.
The third layer exists because the first two produce something keyword research cannot: a named prompt,
a named competitor, and the answer text that proves the loss.
First
The reporting names the losses
Specific prompts where a competitor is recommended and the client is not, with the verbatim answer attached to each one.
Second
The connector closes what it can
Anything answerable from brand data becomes a connector capability rather than a media line. That is the cheapest fix available.
Third
Paid covers the remainder
What is left is a media buy against known questions with a known rival, measured on the same dashboard as everything else.
Advertising inside assistants is early and the pricing is unsettled. That is an argument for being
in it while it is cheap and already instrumented, rather than arriving after the cost has found its level.
BeaconCommercials
The second brand costs a fraction of the first, and Philips already paid for the path.
Establish
One category proves it
The same connector then carries the brand into ChatGPT, Gemini and Claude on the same standard, without a second build.
Extend
Adding is content work
A new category is data and copy against standing infrastructure. Each one launches at a fraction of the first.
Compound
Ride the platform spend
The labs are spending billions advancing these platforms. Every capability they ship reaches the brand at content-update cost rather than rebuild cost.
The platform review path is already walked. That is the part that cannot be compressed with budget,
only by having done it before, which is why the next build is shorter than the Philips one was.
BeaconCommercials
Three phases to stand up, and reporting can be bought on its own.
Fees are set per brand and per market. Numbers for the account in question go here.
Phase 1
Working connector
Brand data structured, connector live on one platform, platform review submitted. Reporting switched on in parallel with historical crawler logs backfilled.
Phase 2
Platform experience
Rich answer experiences designed, submitted and approved. Action board live with the first cards assigned to named owners.
Phase 3
Run
Hosting, upkeep, content updates, monthly reporting, and the paid layer once there is enough data to target it properly.
Reporting can be sold on its own and often should be, because it produces the evidence that
justifies the connector. The connector cannot be sold without reporting, because there would be no way to prove it
worked.
BeaconEVAA Pte. Ltd.
EVAA Pte. Ltd.
1 Raffles Place, Singapore · sc@evaa.sg · NVIDIA Inception member