We build Agents. You build the Future.
● Business offer — Communication & advertisingThe AI search & campaigns agent: optimise SEO, prepare your campaigns
Keyword research, content optimisation, internal linking, preparing paid search and social campaigns: a huge share of a communication team's work is repetitive — without being the value you sell, which is creativity and strategy. Your AI agent absorbs that preparation work. Hosted in France — on local inference or an isolated resource — your content, your brands and your advertisers' data stay under control. The creative direction keeps the decision.
Updated on
Fictional simulation — not a client result
Structured brief ready — for the writer to approve.
⛓ Sourced · your website + your internal product pages
I am preparing a keyword and ad table for review — the project manager decides and launches.
✎ Action · campaign groundwork ready for review — the human decides and publishes
For an organisation established outside the European Union, the law of its own country applies; at its request, we put equivalent safeguards in place (human review before any action, a log of the criteria, information for the people concerned).
For a communication or marketing team, or an agency, a Blue Lemon Agent agent strengthens search and campaign preparation — keyword research, SEO optimisation of content, internal linking, briefs, preparation of paid search and social campaigns. Data hosted in France; the AI model is called at its publisher, except with the Sovereign model option. With that option, your content, your editorial line and your advertisers' data are never exposed to a foreign service, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. The time saved on preparation is reinvested in creativity and strategy. Live within a few weeks.
Reference points describing our offer, not results measured at a client. The scale of the gain is confirmed by a pilot on your own scope.
Indicative estimate, to be confirmed by a pilot on your real files (general terms of sale, article 4).
Three options, one agent
A search & campaigns agent (keywords, SEO optimisation, briefs, paid search preparation), installed and operated for you. Choose according to how you work. Prices exclude VAT — annual subscription, the time it takes for the gains to settle in.
Setup + controlled subscription
- Installation, configuration and training for your teams
- Operation, human oversight, updates and support
- Hosting in France, a dedicated and isolated resource
All inclusive, no setup fee
- Setup included (installation, configuration, training)
- Operation, human oversight, updates and support
- Hosting in France, managed end to end
On site, you own it
- Hardware installed on your premises (you own it)
- French / European AI models run locally
- Secure remote maintenance (Pro support included)
What does the SEO and campaigns agent actually do?
Here are the capabilities included in the agent, then the neighbouring agents that can be added to it. The agent carries out by itself what you authorise it to do; you decide what it does on its own and what it submits to the creative direction, who keeps the decision.
Search & campaigns
Keyword research, SEO optimisation, internal linking and paid search campaign preparation — this agent's main function.
Monitoring & campaign reporting
Aggregating performance, monitoring competitors and producing reports ready to comment on for the advertiser.
Knowledge base & briefs
Centralising guidelines, briefs and history so the right item can be found at the right moment.
Need to go further?
These agents handle a different business process, with their own owner and their own price. They are added to this one.
Copywriting & content creation
Writing articles, pages, newsletters and advertising hooks in your editorial voice, for review.
Writing & content creation from 224 € excl. VAT / month Copywriting & content →Community management
Preparing the editorial calendar, the social media posts and the replies to comments, subject to approval.
Community management from 320 € excl. VAT / month Community management →Visual creation & adaptations
Adapting a visual to the formats of the networks and the banners, respecting the brand guidelines.
Visual creation & adaptations from 227 € excl. VAT / month Visual creation →CRM & follow-up
Keeping the tracking of accounts, opportunities and actions up to date, with no tedious re-keying.
Sales agent (meeting notes, follow-ups, CRM) from 307 € excl. VAT / month CRM & follow-up →In 30 minutes we identify the most relevant agent — without oversizing the project.
See the agent at work
4 everyday situations, taken from those that come up most often. Pick one: the exchange unfolds as it would in your organisation.
A scripted demonstration. These exchanges show how the agent behaves — its sources, its refusals, what it leaves to your teams. Nothing is sent from this page, no model is queried here, and the matters named are fictional. That is precisely what we promise your data. The behaviours shown here — monitoring, automation rules, routing and reminders — are configured with you during deployment, from your tools, your rules and your thresholds. The architecture points named in these exchanges — location, local execution, isolation, encryption, role-based access, logging — are not a guarantee attached to the demonstration: they are those of the architecture set out in your quotation, and verified before commissioning.
The company in this demonstration
Fictional companyBelrive Équipement — outdoor equipment and furniture retailer, fictional company
- Sector
- Outdoor equipment and furniture retail, online and in store. A 340-page website, a catalogue in which some references were discontinued in 2024, and campaigns running on bought keywords.
- Headcount
- 68 employees, 3 of them on the editorial team — they are the ones who wrote, in 2022, 2024 and 2026, the three pages that today cover the same subject — plus a website owner and a technical lead.
- Who is served
- Consumers and tradespeople who arrive through search: 1,240 incoming queries bring the visitors in, and 87 of them are about products or services the company no longer provides, for 2,100 visits a month.
- Order of magnitude
- 340 pages online, 1,240 incoming queries, 212 briefs — 118 editorial, 94 campaign — recovered from six shared folders and two mailboxes, and the campaign history with its bought keywords, budgets and written lessons.
- Tools in place
- The indexing configuration, the code repository and its release log, position and query tracking, inbound links, the campaign accounts, and the editorial charter with its two dated versions — the agent reads what is opened to it, nothing is replaced.
- Who decides
- The website owner approves and publishes: no indexing change is applied without them. The editorial lead designates the main page when several pages cover the same subject; the sales lead receives the queries that are really about the offer; the executive team decides on the 84 bought links and on filing a disavowal.
- Improvement points
- 41 pages still carry the no-indexing instruction left over from the 12 May rebuild — 9% of the site for 28% of the traffic, and the fix is one line per page. 87 queries bring 2,100 visits a month to discontinued products, to be redirected to the current catalogue. 14 subjects are covered by two pages or more, as many audiences to reunite. 84 links from 6 domains await an executive decision. And 41 of the 212 briefs can be dated in an hour.
Belrive Équipement sells online and in store, and a large share of its visitors arrive through search. Twelve weeks of falling traffic had been put down to seasonality: set against the release log, they become explainable, and 28% of the traffic is replayed on 41 pages whose fix is a single line. The agent reads the indexing configuration, positions and queries, inbound links, the release log and the base of 212 briefs. It sets the curves side by side, names the cause, writes the fix and puts it on the table with its figures; the website owner publishes, and that is the only gesture left to them. The exchanges that follow cover one morning of monitoring and the three cases it opened: the 84 bought links, the queries that came looking for a discontinued product, and the measurement the agent commits to instead of a ranking promise.
This company, its figures and the exchanges that follow were invented for the demonstration. They illustrate a common situation; they describe no real client.
Fictional simulation — not a client result
· A no-index instruction has remained on 41 pages since the 12 May redesign. Those pages have dropped out of the results. Traffic had been falling for twelve weeks: the connection is now made, and those twelve weeks become explainable.
· A page ranks on a query you cannot serve — a product you stopped selling in 2024. It receives 340 visits a month.
· Three pages cover the same subject and compete on the same queries.
· A supplier bought 84 links pointing to your site in June. They come from 6 domains. morning-watch_4-flags.pdf4 flags · 41 pages de-indexed
⛓ Source · indexing configuration, release log, incoming queries, inbound links
What I record: the 12 May redesign was prepared on a test environment, where no-indexing is normal. It went live with that setting, and it stayed on 41 pages.
What it produced: the 41 pages left the results within three weeks. Total traffic fell by 28%, and that fall was attributed to seasonality in your monthly reporting.
Why the cause was hard to see: the fall is gradual, it looks like seasonality, and the 41 pages are 9% of the site and 28% of the traffic — that is, your best pages.
What I have prepared: the 41 pages listed, the exact line to remove on each, the fix written on a branch of the repository, and the re-examination request ready to submit once it is live. Twenty minutes of work, not a day.
What stays in your hands: putting it live. An indexing change made in the wrong place de-indexes the rest of the site — that is the kind of act that gets reread and carries a name.
And the estimate, honestly: the pages do not return to their previous positions, and it takes several weeks. On timing and level I commit to the measurement rather than to a promise — you will have the curve, page by page, every week. 41-pages_de-indexed.pdf9% of the site · 28% of traffic · 12 weeks
⛓ Source · configuration of 12/05, 41 pages, 12 weeks of traffic
Routing follows what is being lost: a page out of the index goes to technical and the site owner, within the hour — every day counts; a query you cannot serve to sales, not to the SEO lead, because it is a question of offer; competing pages to the editorial lead; bought links to management, immediately.
With a chase: 4 h on a de-indexing, 24 h on bought links, 7 days on the rest. Then a monthly summary: by page and by query, never by writer.
What that gives you in this one morning: 41 pages ready to come back into the index, the ones that carried 28% of your traffic; 340 monthly visits redirected towards an offer you still sell; three pages that stop fighting over the same queries; and 84 bought links raised with management while there is still time to decide on them.
What you gain from tomorrow: twelve weeks of decline that stop being put down to the season, and a one-line fix per page. The site owner approves and publishes in minutes, from a costed list, page by page. The time taken back from preparation goes to creative work.
What I do commit to: not a position, nobody sells a ranking, but the work that builds one: every page measured, every gap dated, every gain tied to the release that produced it, in a monthly summary.
On what I see: your content, your editorial line and your advertisers' data stay on an isolated resource in France, access by role, logged, withdrawn with a single word. The next step is ready: fifteen minutes to plug in your indexing configuration, and I hand you the 41 lines to remove.
✎ Framework · publishing stays with the site owner — measurement rather than a promise
Fictional simulation — not a client result
What I record: 84 links appeared between 4 and 19 June, from only 6 domains, with very similar anchor texts and on the home pages of sites unrelated to your business.
What it produced: your positions on three queries improved in July. The result is real, and that is what makes this delicate.
What it commits you to: link buying works until the day it is detected. The penalty is retroactive and falls on the site, not only on the bought links, and it is rarely recovered from within the year. That is why the path is open today, and it is written below: the question to the supplier in writing, then the disavowal prepared, checked, ready to file the day you decide.
What I have prepared for the disavowal, without filing it: the file of the 6 domains, and above all the reverse check — the 312 legitimate links those same domains do not explain, so that none of them goes with it. A badly made disavowal removes legitimate links, and it is decided with somebody who knows your history: in practice it is an irreversible act.
What I provide so it can be settled: the 84 links, the 6 domains, the dates, and the supplier's contract if it is on file — it does not mention link buying, it speaks of "visibility".
What I propose, and in this order: the question put to the supplier, in writing — it is drafted — before any technical action. 84-links_6-domains.pdf15 days · real result · file ready for management
⛓ Source · 84 inbound links, 6 domains, supplier contract
· Showing the engine content different from the visitor's. It solves a great deal and it gets detected.
· Creating pages built for a query with no real content — one template rolled out across three hundred towns.
· Buying, exchanging or manufacturing links.
· Repeating a phrase until it is unreadable to hit a density.
· Publishing a review, testimonial or rating that does not exist, including as structured data.
The fifth deserves a word: declaring a 4.8-out-of-5 rating in a page's structured data makes stars appear in the results, and it works. Across your pages I spotted three declaring a rating with no published review to support it — spotted before any check does, and the fix is one line per page.
What I do instead, and it is the real work: writing what the page actually says, in the words people use, and removing what stops the engine reading it.
It is not slower: it is simply less spectacular for the first three months, and what rises stays yours. 5-techniques_3-declared-ratings.pdf3 pages spotted · a one-line fix each
✎ Framework · 5 shortcuts set aside — 3 pages spotted, a one-line fix
Fictional simulation — not a client result
What I record: the page ranks well on a query about a product discontinued in 2024. It receives 340 monthly visits. The immediate bounce rate is 84%, against 41% across the rest of the site.
What that means: 340 people look for something, arrive at your site, and leave. Those are 340 buying intentions a month with no answer yet — and two products in your current catalogue answer them.
The three options, and the first is not technical: offer a replacement — if a current product meets the need, the page says so and the 340 visits become useful; explain the discontinuation and point elsewhere; or let the page decline.
What I matched so the choice rests on figures: two products in your current catalogue cover the discontinued one's need, and visitors leaving that page go on to look at one of the two in 12% of cases — with no link taking them there. The text of the "replacement" page is written, and so is the "discontinuation explained" one.
What I set aside, and why: sending people to a product other than the one they sought is perceived as deception, by the visitor and by the engine — and the page would lose both. The two texts I am handing you do the same work without that drawback.
What I also looked at: across your 1,240 incoming queries, 87 concern products or services you no longer provide, for 2,100 monthly visits — that much demand already coming to you, to redirect towards what you sell today. 87-queries_2100-visits.pdf84% bounce · 3 options
⛓ Source · incoming queries, catalogue, immediate bounce rate
What I record: three pages cover the same subject, written in 2022, 2024 and 2026 by three people. They alternate on the same queries, and none settles.
What that produces: the engine chooses each time, and it does not always choose the best. The 2022 page appears most often — the oldest and the least current.
What I propose: designating one main page, moving into it what the other two contribute, and pointing the other two at it. The merged page is already written for all three cases — one per candidate — and each keeps, line for line, what the other two had that was unique.
What is yours to do, and it is not technical: designating which of the three becomes the main one. That is an editorial decision, and it is signed: a deleted page takes its inbound links with it, and two of the three have some — 41 and 18 respectively, none on the 2026 one.
What I provide so it can be settled: the three pages, what each contains that the other two do not, their inbound links, and their respective positions over twelve months.
Across the whole site: 14 subjects are covered by two or more pages — that many audiences to bring back together, and the same method applies to each. 14-subjects_competing-pages.pdf3 pages · 2 with inbound links
⛓ Source · 3 pages, positions over 12 months, inbound links, 14 duplicated subjects
Fictional simulation — not a client result
What it centralises today: your editorial guidelines and their two successive versions with their dates, 212 briefs — 118 content, 94 campaign — recovered from six shared directories and two mailboxes, the history of your 340 pages with their publication date and their changes, and your campaign history with the keywords bought, the budgets, and the lessons written at the end of each one.
What that fixes immediately, and you lived it this morning: the three pages competing with each other each have a brief in the knowledge base. All three briefs ask for the same subject, written in 2022, 2024 and 2026 by three people who did not have the other two. The duplication was not born in the writing, it was born in the brief — and that is the only way to avoid it next time.
How I use it: before writing a recommendation, I look for the brief and the page's history. A recommendation that ignores the original brief often proposes undoing a decision taken deliberately. Across my 64 recommendations this quarter, 7 would have done exactly that — and I know because the brief said so.
The figure I publish about myself: 41 of the 212 briefs carry no date, and 19 carry no requester's name. I serve them saying so, like an undated document, and alone they support no recommendation. What I propose: two mandatory fields when a brief is filed, date and requester — and the 41 can be dated in an hour if someone reopens them.
⛓ Source · 212 briefs centralised, 340 pages historised, 7 recommendations caught by the brief, 41 undated briefs
Why no guarantee of place exists: ranking depends on what your competitors do, on algorithm changes nobody knows in advance, and on what the engine judges to answer the intent behind the query — an intent that shifts.
What I can say, and it is verifiable: what is stopping a page being read today, what better-ranked pages contain that yours does not, and what you have that they do not.
What I can measure: positions, visits, the queries bringing people to you, and the immediate bounce rate per query — the only indicator that says whether the visit was any use.
The two false successes the report names instead of counting: a rise in position with no rise in visits, and a rise in visits with no effect on what you sell. Both happen, and both are readily presented as successes — you will see them spelled out, and that is what makes the rest of the report believable.
On the 41 pages once restored: you will have the positions recovered and those that will not be. There will be some — named page by page, with what I propose for each. what-is-measured.pdf4 indicators · 2 false successes named
✎ Framework · what is measured — 4 verifiable indicators
What is kept: positions and queries, configuration changes with their date, recommendations made and what became of them, inbound links, and the site's releases.
Why the releases: that is the chart that was missing this morning, and it now sits alongside the traffic. A 28% traffic fall attributed to seasonality would have been explained in ten minutes if the 12 May redesign appeared on the same chart.
What is not kept: no data on individual visitors, no browsing profile, no statistics per writer, and no data on competitor sites beyond what is public.
What the monthly summary contains: pages out of the index, queries you cannot serve, subjects covered by several pages, and recommendations not applied. The last is the one missing almost everywhere else — a recommendation that is never applied is not a recommendation, it is a list. what-is-kept.pdf5 items kept · 4 never collected
✎ Framework · retention periods to be set by the company
Your case is not here? That is exactly what a 30-minute conversation is for. Discuss your project →
The stages of your AI agent project
Audit & scoping
30 minutes to target the use case with the best return.
Quote or direct sign-up
A catalogue offer is bought online; a specific need gets a costed quote.
Design
We design the agent and its guardrails.
Integration & testing
We connect your tools to the agent, which is itself hosted in France.
Rollout
Going live and training your team.
Operation
Continuous supervision and improvement.
Why AI matters to communication — and why it worries people
SEO, content production and campaign management demand a growing volume of work, while advertisers expect ever more responsiveness. But entrusting your content, your editorial line and your clients' data to a consumer tool raises a real question of ownership and confidentiality.
! The issue
A communication team is caught between advertisers who want volume, visibility and measurable results, and an exploding production workload (keywords, content, adaptations, reporting). Yet most consumer AI tools amount to entrusting your content, your editorial strategy, your brands and your advertisers' data to a third party, often hosted outside Europe and subject to the Cloud Act — with the risk that your work feeds the training of a public model.
✓ Our answer
AI is only of interest to communication if it is sovereign and confidential by design. Local inference or an isolated resource hosted in France, systematic human oversight, decisions reserved to the creative direction and the project manager: the time saved on preparation is never paid for in lost intellectual property. The aim is not to replace the creatives, but to give them back time to create and advise.
Ownership of the content & advertisers' data: sovereignty & compliance
A communication team handles high-value content, brands and client data. Here is how the architecture of our agents protects them, campaign by campaign.
Local inference
With the 100% Sovereign package, the agent runs on a machine belonging to the team: no content, no brief and no advertiser data leaves the network.
Hosting in France
Otherwise, with the Agility and Serenity packages, a dedicated, isolated resource hosted in France, under French law: data hosted in France; the AI model is called at its publisher, except with the Sovereign model option.
Reduced extraterritorial exposure
Your content and your brands: architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
Your content stays yours
Your work feeds no public model: the intellectual property of your texts, visuals and brands stays with you.
One isolated resource per account
No pooling: a strictly dedicated environment, ideal for keeping each advertiser's data apart.
AI Act: governed deployment
The agent carries out what you have authorised; no campaign launched automatically; traceability and human oversight from end to end.
What depends on the architecture chosen These points are not general guarantees: they are settled deployment by deployment, in the quotation.
- The applicable location is that of the architecture set out in the quotation and verified before commissioning.
- Local execution is announced only for the configuration explicitly described and accepted in the quotation.
- The applicable isolation depends on the deployment mode set out in the quotation; no dedicated isolation is presumed.
- The events logged, their content, their retention period and who may access them are defined for the deployment chosen.
How much time can a communication team win back?
By automating keyword research, content optimisation and campaign preparation, a team can aim for a reduction by half in preparation time on standardised tasks — reinvested in creativity and strategy.
Four guarantees that matter to a communication team
Related resources
Your questions, our answers
Can AI really improve my search rankings (SEO)?
Is my content used to train a public model?
Who owns the intellectual property of the content produced?
Is my advertisers' data confidential?
Does the agent launch the campaigns on its own?
Do we have to change SEO tools or ad platforms?
How long does it take to deploy an agent?
Other agents for communication & advertising
Let us estimate the potential for your communication
30 minutes to identify the use case with the best return — hosted in France, supervised, with no commitment.